Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1 | // |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2 | // Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include "Utils.hpp" |
| 9 | |
| 10 | #include "ConversionUtils.hpp" |
Matthew Sloyan | 9b088d9 | 2020-09-14 15:12:55 +0100 | [diff] [blame] | 11 | |
| 12 | #include <armnn/utility/NumericCast.hpp> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 13 | #include <armnnUtils/TensorUtils.hpp> |
| 14 | |
| 15 | #include <half/half.hpp> |
| 16 | |
| 17 | using Half = half_float::half; |
| 18 | |
| 19 | namespace armnn_driver |
| 20 | { |
| 21 | |
| 22 | using namespace armnn; |
| 23 | using namespace android::nn; |
| 24 | |
| 25 | template<typename HalPolicy, |
| 26 | typename HalOperation = typename HalPolicy::Operation, |
| 27 | typename HalModel = typename HalPolicy::Model> |
| 28 | bool IsQSymmDequantizeForWeights(const HalOperation& operation, const HalModel& model) |
| 29 | { |
| 30 | using HalOperand = typename HalPolicy::Operand; |
| 31 | using HalOperationType = typename HalPolicy::OperationType; |
| 32 | |
| 33 | const HalOperand* operand = GetInputOperand<HalPolicy>(operation, 0, model); |
| 34 | if (!operand) |
| 35 | { |
| 36 | return false; |
| 37 | } |
| 38 | |
| 39 | if(!IsQSymm8(*operand)) |
| 40 | { |
| 41 | // Only QSymm8 weights are dequantized on the fly by the driver |
| 42 | return false; |
| 43 | } |
| 44 | |
| 45 | if (!IsOperandConstant<HalPolicy>(*operand)) |
| 46 | { |
| 47 | // Non-const input is not accepted for weights |
| 48 | return false; |
| 49 | } |
| 50 | |
| 51 | // Iterate through all the operations and find the operation feeding from the Dequantize output |
| 52 | const size_t outputIndex = operation.outputs[0]; |
| 53 | for (uint32_t operationIdx = 0; operationIdx < getMainModel(model).operations.size(); ++operationIdx) |
| 54 | { |
| 55 | const auto& operationIt = getMainModel(model).operations[operationIdx]; |
| 56 | switch (operationIt.type) |
| 57 | { |
| 58 | case HalOperationType::FULLY_CONNECTED: |
| 59 | if (outputIndex == operationIt.inputs[1]) // Weights are bound to slot 1 |
| 60 | { |
| 61 | // If the output is going into the FC weights return true |
| 62 | return true; |
| 63 | } |
| 64 | break; |
| 65 | case HalOperationType::LSTM: |
| 66 | for (size_t k = 0; k < operationIt.inputs.size(); ++k) |
| 67 | { |
| 68 | if (outputIndex == operationIt.inputs[k]) |
| 69 | { |
| 70 | // If the output is going into the LSTM weights return true |
| 71 | return true; |
| 72 | } |
| 73 | } |
| 74 | break; |
| 75 | default: |
| 76 | break; |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | return false; |
| 81 | } |
| 82 | |
| 83 | template<typename HalPolicy, |
| 84 | typename HalOperation = typename HalPolicy::Operation, |
| 85 | typename HalModel = typename HalPolicy::Model> |
| 86 | bool SetupAndTrackLayerOutputSlotAndOverrideTensorInfo(const HalOperation& operation, |
| 87 | uint32_t operationOutputIndex, |
| 88 | armnn::IConnectableLayer& layer, |
| 89 | uint32_t layerOutputIndex, |
| 90 | const HalModel& model, |
| 91 | ConversionData& data, |
| 92 | const armnn::TensorInfo tensor_info) |
| 93 | { |
| 94 | using HalOperand = typename HalPolicy::Operand; |
| 95 | |
| 96 | const HalOperand* outputOperand = GetOutputOperand<HalPolicy>(operation, operationOutputIndex, model); |
| 97 | if ((outputOperand == nullptr) || (operationOutputIndex >= layer.GetNumOutputSlots())) |
| 98 | { |
| 99 | return false; |
| 100 | } |
| 101 | |
| 102 | armnn::IOutputSlot& outputSlot = layer.GetOutputSlot(layerOutputIndex); |
| 103 | |
| 104 | const uint32_t operandIndex = operation.outputs[operationOutputIndex]; |
| 105 | data.m_OutputSlotForOperand[operandIndex] = &outputSlot; |
| 106 | |
| 107 | outputSlot.SetTensorInfo(tensor_info); |
| 108 | |
| 109 | return true; |
| 110 | } |
| 111 | |
| 112 | template<typename HalPolicy, |
Sadik Armagan | 92b5fd1 | 2021-04-26 09:52:06 +0100 | [diff] [blame] | 113 | typename HalOperation = typename HalPolicy::Operation, |
| 114 | typename HalModel = typename HalPolicy::Model> |
| 115 | bool ConvertCast(const HalOperation& operation, |
| 116 | const HalModel& model, |
| 117 | ConversionData& data) |
| 118 | { |
| 119 | using HalOperand = typename HalPolicy::Operand; |
| 120 | |
| 121 | ALOGV("HalPolicy::ConvertCast()"); |
| 122 | |
| 123 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 124 | |
| 125 | if (!input.IsValid()) |
| 126 | { |
| 127 | return Fail("%s: Operation has invalid inputs", __func__); |
| 128 | } |
| 129 | |
| 130 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 131 | if (!output) |
| 132 | { |
| 133 | return Fail("%s: Could not read output 0", __func__); |
| 134 | } |
| 135 | |
| 136 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 137 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 138 | |
| 139 | bool isSupported = false; |
| 140 | |
| 141 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 142 | { |
| 143 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 144 | IsCastSupported, |
| 145 | data.m_Backends, |
| 146 | isSupported, |
| 147 | inputInfo, |
| 148 | outputInfo); |
| 149 | }; |
| 150 | |
| 151 | if(!IsDynamicTensor(outputInfo)) |
| 152 | { |
| 153 | validateFunc(outputInfo, isSupported); |
| 154 | } |
| 155 | else |
| 156 | { |
| 157 | isSupported = AreDynamicTensorsSupported(); |
| 158 | } |
| 159 | |
| 160 | if (!isSupported) |
| 161 | { |
| 162 | return false; |
| 163 | } |
| 164 | |
| 165 | IConnectableLayer* layer = data.m_Network->AddCastLayer(); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 166 | if (!layer) |
| 167 | { |
| 168 | return Fail("%s: Could not add the CastLayer", __func__); |
| 169 | } |
Sadik Armagan | 92b5fd1 | 2021-04-26 09:52:06 +0100 | [diff] [blame] | 170 | input.Connect(layer->GetInputSlot(0)); |
| 171 | |
| 172 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 173 | } |
| 174 | |
| 175 | template<typename HalPolicy, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 176 | typename HalOperation = typename HalPolicy::Operation, |
| 177 | typename HalModel = typename HalPolicy::Model> |
Teresa Charlin | 7f5b51e | 2021-09-01 14:19:38 +0100 | [diff] [blame] | 178 | bool ConvertChannelShuffle(const HalOperation& operation, |
| 179 | const HalModel& model, |
| 180 | ConversionData& data) |
| 181 | { |
| 182 | using HalOperand = typename HalPolicy::Operand; |
| 183 | using HalOperandType = typename HalPolicy::OperandType; |
| 184 | |
| 185 | ALOGV("HalPolicy::ConvertChannelShuffle()"); |
| 186 | |
| 187 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 188 | if (!input.IsValid()) |
| 189 | { |
| 190 | return Fail("%s: Operation has invalid inputs", __func__); |
| 191 | } |
| 192 | auto inputDimensions = static_cast<int32_t>(input.GetTensorInfo().GetNumDimensions()); |
| 193 | |
| 194 | ChannelShuffleDescriptor descriptor; |
| 195 | |
| 196 | int32_t groups; |
| 197 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, groups, model, data)) |
| 198 | { |
| 199 | return Fail("%s: Operation has invalid or unsupported number of groups operand", __func__); |
| 200 | } |
| 201 | descriptor.m_NumGroups = static_cast<uint32_t>(groups); |
| 202 | |
| 203 | int32_t axis; |
| 204 | if (!GetInputScalar<HalPolicy>(operation, 2, HalOperandType::INT32, axis, model, data)) |
| 205 | { |
| 206 | return Fail("%s: Operation has invalid or unsupported dimension channel shuffle operand", __func__); |
| 207 | } |
| 208 | if (((axis < -inputDimensions) && (axis < 0)) || ((axis >= inputDimensions) && (axis > 0))) |
| 209 | { |
| 210 | return Fail("%s: Operation has invalid dimension: %d. It is out of bounds [-%d, %d))", __func__, axis, |
| 211 | inputDimensions, inputDimensions); |
| 212 | } |
| 213 | int positiveAxis = (axis < 0) ? inputDimensions + axis : axis; |
| 214 | descriptor.m_Axis = static_cast<uint32_t>(positiveAxis); |
| 215 | |
| 216 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 217 | if (!output) |
| 218 | { |
| 219 | return Fail("%s: Could not read output 0", __func__); |
| 220 | } |
| 221 | |
| 222 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 223 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 224 | |
| 225 | bool isSupported = false; |
| 226 | |
| 227 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 228 | { |
| 229 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 230 | IsChannelShuffleSupported, |
| 231 | data.m_Backends, |
| 232 | isSupported, |
| 233 | inputInfo, |
| 234 | outputInfo, |
| 235 | descriptor); |
| 236 | }; |
| 237 | |
| 238 | if(!IsDynamicTensor(outputInfo)) |
| 239 | { |
| 240 | validateFunc(outputInfo, isSupported); |
| 241 | } |
| 242 | else |
| 243 | { |
| 244 | isSupported = AreDynamicTensorsSupported(); |
| 245 | } |
| 246 | |
| 247 | if (!isSupported) |
| 248 | { |
| 249 | return false; |
| 250 | } |
| 251 | |
| 252 | IConnectableLayer* layer = data.m_Network->AddChannelShuffleLayer(descriptor); |
| 253 | assert(layer != nullptr); |
| 254 | input.Connect(layer->GetInputSlot(0)); |
| 255 | |
| 256 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 257 | } |
| 258 | |
| 259 | template<typename HalPolicy, |
| 260 | typename HalOperation = typename HalPolicy::Operation, |
| 261 | typename HalModel = typename HalPolicy::Model> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 262 | bool ConvertComparison_1_2(const HalOperation& operation, |
| 263 | const HalModel& model, |
| 264 | ConversionData& data, |
| 265 | ComparisonOperation comparisonOperation) |
| 266 | { |
| 267 | using HalOperand = typename HalPolicy::Operand; |
| 268 | |
| 269 | ALOGV("HalPolicy::ConvertComparison()"); |
| 270 | ALOGV("comparisonOperation = %s", GetComparisonOperationAsCString(comparisonOperation)); |
| 271 | |
| 272 | LayerInputHandle input0 = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 273 | LayerInputHandle input1 = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 274 | |
| 275 | if (!(input0.IsValid() && input1.IsValid())) |
| 276 | { |
| 277 | return Fail("%s: Operation has invalid inputs", __func__); |
| 278 | } |
| 279 | |
| 280 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 281 | if (!output) |
| 282 | { |
| 283 | return Fail("%s: Could not read output 0", __func__); |
| 284 | } |
| 285 | |
| 286 | const TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 287 | const TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 288 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 289 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 290 | ComparisonDescriptor descriptor(comparisonOperation); |
| 291 | |
| 292 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 293 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 294 | { |
| 295 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 296 | IsComparisonSupported, |
| 297 | data.m_Backends, |
| 298 | isSupported, |
| 299 | inputInfo0, |
| 300 | inputInfo1, |
| 301 | outputInfo, |
| 302 | descriptor); |
| 303 | |
| 304 | }; |
| 305 | |
| 306 | if(!IsDynamicTensor(outputInfo)) |
| 307 | { |
| 308 | validateFunc(outputInfo, isSupported); |
| 309 | } |
| 310 | else |
| 311 | { |
| 312 | isSupported = AreDynamicTensorsSupported(); |
| 313 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 314 | |
| 315 | if (!isSupported) |
| 316 | { |
| 317 | return false; |
| 318 | } |
| 319 | |
| 320 | IConnectableLayer* layer = data.m_Network->AddComparisonLayer(descriptor); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 321 | if (!layer) |
| 322 | { |
| 323 | return Fail("%s: Could not add the ComparisonLayer", __func__); |
| 324 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 325 | |
| 326 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 327 | if (!isReshapeSupported) |
| 328 | { |
| 329 | return false; |
| 330 | } |
| 331 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 332 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 333 | } |
| 334 | |
| 335 | template<typename HalPolicy, |
| 336 | typename HalOperation = typename HalPolicy::Operation, |
| 337 | typename HalModel = typename HalPolicy::Model> |
| 338 | bool ConvertConv2d_1_2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 339 | { |
| 340 | |
| 341 | using HalOperand = typename HalPolicy::Operand; |
| 342 | using HalOperandType = typename HalPolicy::OperandType; |
| 343 | |
| 344 | ALOGV("HalPolicy::ConvertConv2d_1_2()"); |
| 345 | |
| 346 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 347 | if (!input.IsValid()) |
| 348 | { |
| 349 | return Fail("%s: Operation has invalid inputs", __func__); |
| 350 | } |
| 351 | |
| 352 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 353 | if (!output) |
| 354 | { |
| 355 | return Fail("%s: Could not read output 0", __func__); |
| 356 | } |
| 357 | |
| 358 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 359 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 360 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 361 | Convolution2dDescriptor desc; |
| 362 | desc.m_DataLayout = DataLayout::NHWC; |
| 363 | |
| 364 | // Determine whether padding is implicit or explicit |
| 365 | bool implicitPadding = operation.inputs.size() == 7 || |
| 366 | (operation.inputs.size() >= 8 && |
| 367 | GetInputOperand<HalPolicy>(operation, 7, model)->type == HalOperandType::BOOL); |
| 368 | |
| 369 | if (implicitPadding) |
| 370 | { |
| 371 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 7, model, data); |
| 372 | } |
| 373 | else if (operation.inputs.size() >= 10) |
| 374 | { |
| 375 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 10, model, data); |
| 376 | } |
| 377 | |
| 378 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
| 379 | |
| 380 | // ArmNN does not currently support non-fixed weights or bias |
| 381 | // The NNAPI filter is always OHWI [depth_out, filter_height, filter_width, depth_in] but ArmNN expects the |
| 382 | // filter's height and width indices to match the input's height and width indices so we permute it to OIHW if |
| 383 | // the DataLayout is NCHW |
| 384 | const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ? |
| 385 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, |
| 386 | model, data, OHWIToOIHW) : |
| 387 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 388 | const ConstTensorPin biasPin = |
| 389 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 390 | |
| 391 | if (!weightsPin.IsValid()) |
| 392 | { |
| 393 | return Fail("%s: Operation has invalid weights", __func__); |
| 394 | } |
| 395 | |
| 396 | if (!biasPin.IsValid()) |
| 397 | { |
| 398 | return Fail("%s: Operation has invalid biases", __func__); |
| 399 | } |
| 400 | |
| 401 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 402 | ConstTensor bias = biasPin.GetConstTensor(); |
| 403 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 404 | |
| 405 | ActivationFn activation; |
| 406 | |
| 407 | if (implicitPadding) |
| 408 | { |
| 409 | android::nn::PaddingScheme paddingScheme; |
| 410 | if (!GetInputPaddingScheme<HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 411 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 412 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 413 | !GetInputActivationFunction<HalPolicy>(operation, 6, activation, model, data) || |
| 414 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 8, desc, model, data)) |
| 415 | { |
| 416 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 417 | } |
| 418 | |
| 419 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 420 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 421 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 422 | const uint32_t kernelX = weights.GetShape()[widthIndex]; |
| 423 | const uint32_t kernelY = weights.GetShape()[heightIndex]; |
| 424 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 425 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 426 | |
| 427 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 428 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 429 | |
| 430 | } |
| 431 | else if (operation.inputs.size() >= 10) |
| 432 | { |
| 433 | // explicit padding |
| 434 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 435 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 436 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 437 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 438 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 439 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 440 | !GetInputActivationFunction<HalPolicy>(operation, 9, activation, model, data) || |
| 441 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 11, desc, model, data)) |
| 442 | { |
| 443 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 444 | } |
| 445 | } |
| 446 | else |
| 447 | { |
| 448 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 449 | } |
| 450 | |
| 451 | desc.m_BiasEnabled = true; |
| 452 | Optional<TensorInfo> biases(bias.GetInfo()); |
| 453 | |
| 454 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 455 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 456 | { |
| 457 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 458 | IsConvolution2dSupported, |
| 459 | data.m_Backends, |
| 460 | isSupported, |
| 461 | inputInfo, |
| 462 | outputInfo, |
| 463 | desc, |
| 464 | weights.GetInfo(), |
| 465 | biases); |
| 466 | }; |
| 467 | |
| 468 | if(!IsDynamicTensor(outputInfo)) |
| 469 | { |
| 470 | validateFunc(outputInfo, isSupported); |
| 471 | } |
| 472 | else |
| 473 | { |
| 474 | isSupported = AreDynamicTensorsSupported(); |
| 475 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 476 | |
| 477 | if (!isSupported) |
| 478 | { |
| 479 | return false; |
| 480 | } |
| 481 | |
| 482 | IConnectableLayer* startLayer = |
| 483 | data.m_Network->AddConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
| 484 | |
| 485 | if (!startLayer) |
| 486 | { |
| 487 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 488 | } |
| 489 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 490 | input.Connect(startLayer->GetInputSlot(0)); |
| 491 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 492 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model, |
| 493 | data, nullptr, validateFunc, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 494 | } |
| 495 | |
| 496 | template<typename HalPolicy, |
| 497 | typename HalOperation = typename HalPolicy::Operation, |
| 498 | typename HalModel = typename HalPolicy::Model> |
| 499 | bool ConvertDepthwiseConv2d_1_2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 500 | { |
| 501 | using HalOperand = typename HalPolicy::Operand; |
| 502 | using HalOperandType = typename HalPolicy::OperandType; |
| 503 | |
| 504 | ALOGV("HalPolicy::ConvertDepthwiseConv2d_1_2()"); |
| 505 | |
| 506 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 507 | |
| 508 | if (!input.IsValid()) |
| 509 | { |
| 510 | return Fail("%s: Operation has invalid inputs", __func__); |
| 511 | } |
| 512 | |
| 513 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 514 | |
| 515 | if (!output) |
| 516 | { |
| 517 | return Fail("%s: Could not read output 0", __func__); |
| 518 | } |
| 519 | |
| 520 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 521 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 522 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 523 | // ArmNN does not currently support non-fixed weights or bias |
| 524 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 525 | const HalOperand* weightsOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
| 526 | |
| 527 | if (weightsOperand == nullptr) |
| 528 | { |
| 529 | return Fail("%s: Operand is invalid", __func__); |
| 530 | } |
| 531 | if ( weightsOperand->dimensions[0] != 1) |
| 532 | { |
| 533 | return Fail("%s: Invalid weights; for depthwise convolution, dimension 0 must be 1 but it is %i", |
| 534 | __func__, weightsOperand->dimensions[0] ); |
| 535 | } |
| 536 | |
| 537 | DepthwiseConvolution2dDescriptor desc; |
| 538 | desc.m_DataLayout = DataLayout::NHWC; |
| 539 | |
| 540 | // Determine whether padding is implicit or explicit |
| 541 | bool implicitPadding = operation.inputs.size() == 8 || |
| 542 | (operation.inputs.size() >= 9 && |
| 543 | GetInputOperand<HalPolicy>(operation, 8, model)->type == HalOperandType::BOOL); |
| 544 | |
| 545 | // Look ahead to find the optional DataLayout, if present |
| 546 | const uint32_t dataLayoutFlagIndex = implicitPadding ? 8 : 11; |
| 547 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, dataLayoutFlagIndex, model, data); |
| 548 | |
| 549 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 550 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 551 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 552 | |
Jan Eilers | a20d2b8 | 2021-04-27 09:21:08 +0100 | [diff] [blame] | 553 | // The layout for weights in depthwise is [ 1, H, W, O] and it's the same in ArmNN. No need to permute anything. |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 554 | const ConstTensorPin weightsPin = |
| 555 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 556 | 1, |
| 557 | model, |
Jan Eilers | a20d2b8 | 2021-04-27 09:21:08 +0100 | [diff] [blame] | 558 | data); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 559 | |
| 560 | // Bias is a 1D tensor |
| 561 | const ConstTensorPin biasPin = |
| 562 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 563 | |
| 564 | if (!weightsPin.IsValid()) |
| 565 | { |
| 566 | return Fail("%s: Operation has invalid weights", __func__); |
| 567 | } |
| 568 | |
| 569 | if (!biasPin.IsValid()) |
| 570 | { |
| 571 | return Fail("%s: Operation has invalid biases", __func__); |
| 572 | } |
| 573 | |
| 574 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 575 | ConstTensor bias = biasPin.GetConstTensor(); |
| 576 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 577 | |
| 578 | ActivationFn activation; |
| 579 | |
| 580 | if (implicitPadding) |
| 581 | { |
| 582 | android::nn::PaddingScheme paddingScheme; |
| 583 | if (!GetInputPaddingScheme<HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 584 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 585 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 586 | !GetInputActivationFunction<HalPolicy>(operation, 7, activation, model, data) || |
| 587 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 9, desc, model, data)) |
| 588 | { |
| 589 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 590 | } |
| 591 | |
Jan Eilers | a20d2b8 | 2021-04-27 09:21:08 +0100 | [diff] [blame] | 592 | const uint32_t kernelX = weights.GetShape()[2]; |
| 593 | const uint32_t kernelY = weights.GetShape()[1]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 594 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 595 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 596 | |
| 597 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 598 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 599 | } |
| 600 | else if (operation.inputs.size() >= 11) |
| 601 | { |
| 602 | // explicit padding |
| 603 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 604 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 605 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 606 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 607 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 608 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 609 | !GetInputActivationFunction<HalPolicy>(operation, 10, activation, model, data) || |
| 610 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 12, desc, model, data)) |
| 611 | { |
| 612 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 613 | } |
| 614 | } |
| 615 | else |
| 616 | { |
| 617 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 618 | } |
| 619 | |
| 620 | desc.m_BiasEnabled = true; |
| 621 | Optional<TensorInfo> biases(bias.GetInfo()); |
| 622 | |
| 623 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 624 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 625 | { |
| 626 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 627 | IsDepthwiseConvolutionSupported, |
| 628 | data.m_Backends, |
| 629 | isSupported, |
| 630 | inputInfo, |
| 631 | outputInfo, |
| 632 | desc, |
| 633 | weights.GetInfo(), |
| 634 | biases); |
| 635 | }; |
| 636 | |
| 637 | if(!IsDynamicTensor(outputInfo)) |
| 638 | { |
| 639 | validateFunc(outputInfo, isSupported); |
| 640 | } |
| 641 | else |
| 642 | { |
| 643 | isSupported = AreDynamicTensorsSupported(); |
| 644 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 645 | |
| 646 | if (!isSupported) |
| 647 | { |
| 648 | return false; |
| 649 | } |
| 650 | |
| 651 | IConnectableLayer* startLayer = |
| 652 | data.m_Network->AddDepthwiseConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
| 653 | |
| 654 | if (!startLayer) |
| 655 | { |
| 656 | return Fail("%s: AddDepthwiseConvolution2dLayer failed", __func__); |
| 657 | } |
| 658 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 659 | input.Connect(startLayer->GetInputSlot(0)); |
| 660 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 661 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model, |
| 662 | data, nullptr, validateFunc, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 663 | } |
| 664 | |
| 665 | template<typename HalPolicy, |
| 666 | typename HalOperation = typename HalPolicy::Operation, |
| 667 | typename HalModel = typename HalPolicy::Model> |
| 668 | bool ConvertDequantize_1_2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 669 | { |
| 670 | ALOGV("HalPolicy::ConvertDequantize()"); |
| 671 | |
| 672 | if (IsQSymmDequantizeForWeights<HalPolicy>(operation, model)) |
| 673 | { |
| 674 | // NOTE: QSymm8 weights are dequantized internally by the driver, |
| 675 | // therefore this type of Dequantize is implicitly supported |
| 676 | return true; |
| 677 | } |
| 678 | |
| 679 | return ::ConvertDequantize<HalPolicy>(operation, model, data); |
| 680 | } |
| 681 | |
| 682 | template<typename HalPolicy, |
| 683 | typename HalOperation = typename HalPolicy::Operation, |
| 684 | typename HalModel = typename HalPolicy::Model> |
| 685 | bool ConvertElementwiseUnary(const HalOperation& operation, |
| 686 | const HalModel& model, |
| 687 | ConversionData& data, |
| 688 | UnaryOperation unaryOperation) |
| 689 | { |
| 690 | using HalOperand = typename HalPolicy::Operand; |
| 691 | |
| 692 | ALOGV("HalPolicy::ConvertElementwiseUnary()"); |
| 693 | ALOGV("unaryOperation = %s", GetUnaryOperationAsCString(unaryOperation)); |
| 694 | |
| 695 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 696 | |
| 697 | if (!input.IsValid()) |
| 698 | { |
| 699 | return Fail("%s: Operation has invalid input", __func__); |
| 700 | } |
| 701 | |
| 702 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 703 | if (!output) |
| 704 | { |
| 705 | return Fail("%s: Could not read output 0", __func__); |
| 706 | } |
| 707 | |
| 708 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 709 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 710 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 711 | ElementwiseUnaryDescriptor descriptor(unaryOperation); |
| 712 | |
| 713 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 714 | |
| 715 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 716 | { |
| 717 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 718 | IsElementwiseUnarySupported, |
| 719 | data.m_Backends, |
| 720 | isSupported, |
| 721 | inputInfo, |
| 722 | outputInfo, |
| 723 | descriptor); |
| 724 | }; |
| 725 | |
| 726 | if(!IsDynamicTensor(outputInfo)) |
| 727 | { |
| 728 | validateFunc(outputInfo, isSupported); |
| 729 | } |
| 730 | else |
| 731 | { |
| 732 | isSupported = AreDynamicTensorsSupported(); |
| 733 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 734 | |
| 735 | if (!isSupported) |
| 736 | { |
| 737 | return false; |
| 738 | } |
| 739 | |
| 740 | IConnectableLayer* layer = data.m_Network->AddElementwiseUnaryLayer(descriptor); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 741 | if (!layer) |
| 742 | { |
| 743 | return Fail("%s: Could not add the ElementwiseUnaryLayer", __func__); |
| 744 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 745 | input.Connect(layer->GetInputSlot(0)); |
| 746 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 747 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 748 | } |
| 749 | |
| 750 | template<typename HalPolicy, |
| 751 | typename HalOperation = typename HalPolicy::Operation, |
| 752 | typename HalModel = typename HalPolicy::Model> |
| 753 | bool ConvertExpandDims(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 754 | { |
| 755 | using HalOperand = typename HalPolicy::Operand; |
| 756 | using HalOperandType = typename HalPolicy::OperandType; |
| 757 | |
| 758 | ALOGV("HalPolicy::ConvertExpandDims()"); |
| 759 | |
| 760 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 761 | |
| 762 | if (!input.IsValid()) |
| 763 | { |
| 764 | return Fail("%s: Operation has invalid input", __func__); |
| 765 | } |
| 766 | |
| 767 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 768 | if (!output) |
| 769 | { |
| 770 | return Fail("%s: Operation has invalid output", __func__); |
| 771 | } |
| 772 | |
| 773 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 774 | |
| 775 | int32_t axis; |
| 776 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, axis, model, data)) |
| 777 | { |
| 778 | return Fail("%s: failed to get axis input value", __func__); |
| 779 | } |
| 780 | |
| 781 | TensorShape targetShape; |
| 782 | |
| 783 | try |
| 784 | { |
| 785 | targetShape = armnnUtils::ExpandDims(input.GetTensorInfo().GetShape(), axis); |
| 786 | } |
| 787 | catch (const std::exception& e) |
| 788 | { |
| 789 | return Fail("%s: %s", __func__, e.what()); |
| 790 | } |
| 791 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 792 | ReshapeDescriptor reshapeDescriptor; |
| 793 | reshapeDescriptor.m_TargetShape = targetShape; |
| 794 | |
| 795 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 796 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 797 | { |
| 798 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 799 | IsReshapeSupported, |
| 800 | data.m_Backends, |
| 801 | isSupported, |
| 802 | input.GetTensorInfo(), |
| 803 | outputInfo, |
| 804 | reshapeDescriptor); |
| 805 | }; |
| 806 | |
| 807 | if(!IsDynamicTensor(outputInfo)) |
| 808 | { |
Nikhil Raj | c5e0bb0 | 2021-04-02 10:02:16 +0100 | [diff] [blame] | 809 | if (targetShape != outputInfo.GetShape()) |
| 810 | { |
| 811 | return Fail("%s: Shape of the output operand does not match the resolved expanded shape", __func__); |
| 812 | } |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 813 | validateFunc(outputInfo, isSupported); |
| 814 | } |
| 815 | else |
| 816 | { |
| 817 | isSupported = AreDynamicTensorsSupported(); |
| 818 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 819 | |
| 820 | if (!isSupported) |
| 821 | { |
| 822 | return false; |
| 823 | } |
| 824 | |
| 825 | IConnectableLayer* layer = data.m_Network->AddReshapeLayer(reshapeDescriptor); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 826 | if (!layer) |
| 827 | { |
| 828 | return Fail("%s: Could not add the ReshapeLayer", __func__); |
| 829 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 830 | input.Connect(layer->GetInputSlot(0)); |
| 831 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 832 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 833 | } |
| 834 | |
| 835 | template<typename HalPolicy, |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 836 | typename HalOperation = typename HalPolicy::Operation, |
| 837 | typename HalModel = typename HalPolicy::Model> |
| 838 | bool ConvertGather(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 839 | { |
| 840 | using HalOperand = typename HalPolicy::Operand; |
| 841 | using HalOperandType = typename HalPolicy::OperandType; |
| 842 | |
| 843 | ALOGV("HalPolicy::ConvertGather()"); |
| 844 | |
| 845 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 846 | if (!input.IsValid()) |
| 847 | { |
| 848 | return Fail("%s: Operation has invalid input", __func__); |
| 849 | } |
| 850 | auto inputDimensions = input.GetTensorInfo().GetNumDimensions(); |
| 851 | |
| 852 | LayerInputHandle indices = ConvertToLayerInputHandle<HalPolicy>(operation, 2, model, data); |
| 853 | if (!indices.IsValid()) |
| 854 | { |
| 855 | return Fail("%s: Operation has invalid indices", __func__); |
| 856 | } |
| 857 | auto indicesDimensions = indices.GetTensorInfo().GetNumDimensions(); |
| 858 | |
| 859 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 860 | if (!output) |
| 861 | { |
| 862 | return Fail("%s: Operation has invalid output", __func__); |
| 863 | } |
| 864 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 865 | auto outputDimensions = outputInfo.GetNumDimensions(); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 866 | if (outputDimensions != inputDimensions + indicesDimensions - 1) |
| 867 | { |
| 868 | return Fail("%s: Operation has invalid output dimensions: %d. Output must be an (%d + %d - 1)-D tensor", |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 869 | __func__, outputDimensions, inputDimensions, indicesDimensions); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 870 | } |
| 871 | |
Finn Williams | f769f29 | 2021-06-25 12:53:09 +0100 | [diff] [blame] | 872 | uint32_t axis; |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 873 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, axis, model, data)) |
| 874 | { |
| 875 | return Fail("%s: Operation has invalid or unsupported axis operand", __func__); |
| 876 | } |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 877 | if (((axis < -inputDimensions) && (axis < 0)) || ((axis >= inputDimensions) && (axis > 0))) |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 878 | { |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 879 | return Fail("%s: Operation has invalid axis: %d. It is out of bounds [-%d, %d))", __func__, axis, |
| 880 | inputDimensions, inputDimensions); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 881 | } |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 882 | |
| 883 | GatherDescriptor desc; |
| 884 | desc.m_Axis = axis; |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 885 | |
| 886 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 887 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 888 | { |
| 889 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 890 | IsGatherSupported, |
| 891 | data.m_Backends, |
| 892 | isSupported, |
| 893 | input.GetTensorInfo(), |
| 894 | indices.GetTensorInfo(), |
| 895 | outputInfo, |
| 896 | desc); |
| 897 | }; |
| 898 | |
| 899 | if(!IsDynamicTensor(outputInfo)) |
| 900 | { |
| 901 | validateFunc(outputInfo, isSupported); |
| 902 | } |
| 903 | else |
| 904 | { |
| 905 | isSupported = AreDynamicTensorsSupported(); |
| 906 | } |
| 907 | |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 908 | if (!isSupported) |
| 909 | { |
| 910 | return false; |
| 911 | } |
| 912 | |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 913 | IConnectableLayer* layer = data.m_Network->AddGatherLayer(desc); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 914 | if (!layer) |
| 915 | { |
| 916 | return Fail("%s: Could not add the GatherLayer", __func__); |
| 917 | } |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 918 | input.Connect(layer->GetInputSlot(0)); |
| 919 | indices.Connect(layer->GetInputSlot(1)); |
| 920 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 921 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 922 | } |
| 923 | |
| 924 | template<typename HalPolicy, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 925 | typename HalOperation = typename HalPolicy::Operation, |
| 926 | typename HalModel = typename HalPolicy::Model> |
| 927 | bool ConvertGroupedConv2d(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 928 | { |
| 929 | using HalOperand = typename HalPolicy::Operand; |
| 930 | using HalOperandType = typename HalPolicy::OperandType; |
| 931 | |
| 932 | ALOGV("HalPolicy::ConvertGroupedConv2d()"); |
| 933 | |
| 934 | // |
| 935 | // Parse data |
| 936 | // |
| 937 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 938 | if (!input.IsValid()) |
| 939 | { |
| 940 | return Fail("%s: Operation has invalid inputs", __func__); |
| 941 | } |
| 942 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 943 | |
| 944 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 945 | if (!output) |
| 946 | { |
| 947 | return Fail("%s: Could not read output 0", __func__); |
| 948 | } |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 949 | TensorInfo outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 950 | |
| 951 | // Look ahead to determine data layout |
| 952 | DataLayout dataLayout = DataLayout::NHWC; |
| 953 | if (operation.inputs.size() == 12) |
| 954 | { |
| 955 | dataLayout = OptionalDataLayout<HalPolicy>(operation, 11, model, data); |
| 956 | } |
| 957 | else |
| 958 | { |
| 959 | dataLayout = OptionalDataLayout<HalPolicy>(operation, 8, model, data); |
| 960 | } |
| 961 | |
| 962 | // NOTE: |
| 963 | // NNAPI weights are always OHWI, i.e. [depth_out, filter_height, filter_width, depth_group], |
| 964 | // but Arm NN expects the filter's height and width indices to match the input's height and |
| 965 | // width indices so when the DataLayout is NCHW, we need to permute the weights to OIHW |
| 966 | const PermutationVector ohwiToOihw = { 0u, 2u, 3u, 1u }; |
| 967 | const ConstTensorPin weightsPin = (dataLayout == DataLayout::NCHW) ? |
| 968 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, |
| 969 | model, data, ohwiToOihw) : |
| 970 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 971 | const ConstTensorPin biasesPin = |
| 972 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 973 | if (!weightsPin.IsValid() || !biasesPin.IsValid()) |
| 974 | { |
| 975 | return Fail("%s: Operation has invalid inputs", __func__); |
| 976 | } |
| 977 | |
| 978 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 979 | ConstTensor biases = biasesPin.GetConstTensor(); |
| 980 | SanitizeBiasQuantizationScale(biases.GetInfo(), weights.GetInfo(), inputInfo); |
| 981 | |
| 982 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 983 | const TensorShape& outputShape = outputInfo.GetShape(); |
| 984 | const TensorShape& weightsShape = weights.GetShape(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 985 | |
| 986 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| 987 | const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| 988 | const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 989 | const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 990 | |
| 991 | Convolution2dDescriptor desc; |
| 992 | desc.m_DataLayout = dataLayout; |
| 993 | desc.m_BiasEnabled = true; |
| 994 | |
Finn Williams | f769f29 | 2021-06-25 12:53:09 +0100 | [diff] [blame] | 995 | unsigned int numGroups; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 996 | ActivationFn activation; |
| 997 | |
| 998 | if (operation.inputs.size() == 12) |
| 999 | { |
| 1000 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 1001 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 1002 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 1003 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 1004 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 1005 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 1006 | !GetInputScalar<HalPolicy>(operation, 9, HalOperandType::INT32, numGroups, model, data) || |
| 1007 | !GetInputActivationFunction<HalPolicy>(operation, 10, activation, model, data)) |
| 1008 | { |
| 1009 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 1010 | } |
| 1011 | |
| 1012 | } |
| 1013 | else if (operation.inputs.size() == 9) |
| 1014 | { |
| 1015 | android::nn::PaddingScheme paddingScheme; |
| 1016 | if (!GetInputPaddingScheme<HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 1017 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 1018 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 1019 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, numGroups, model, data) || |
| 1020 | !GetInputActivationFunction<HalPolicy>(operation, 7, activation, model, data)) |
| 1021 | { |
| 1022 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 1023 | } |
| 1024 | |
| 1025 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 1026 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 1027 | |
| 1028 | const uint32_t kernelX = weightsShape[widthIndex]; |
| 1029 | const uint32_t kernelY = weightsShape[heightIndex]; |
| 1030 | |
| 1031 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 1032 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 1033 | } |
| 1034 | else |
| 1035 | { |
| 1036 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 1037 | } |
| 1038 | |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1039 | // Equivalent to outputShape[channelsIndex], but we can't know the outputShape in the case of dynamic tensors |
| 1040 | const unsigned int outputChannels = weightsShape[0]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1041 | |
| 1042 | const unsigned int channelsPerGroup = weightsShape[channelsIndex]; |
| 1043 | const unsigned int channelMultiplier = outputChannels / numGroups; |
| 1044 | |
| 1045 | // |
| 1046 | // Validate all relevant inputs |
| 1047 | // |
| 1048 | if (numGroups <= 0) |
| 1049 | { |
| 1050 | return Fail("%s: Number of groups must be greater than 0. Got: %d", __func__, numGroups); |
| 1051 | } |
| 1052 | |
| 1053 | if (outputChannels % numGroups != 0u) |
| 1054 | { |
| 1055 | return Fail("%s: Output channels must be divisible by the number of groups", __func__); |
| 1056 | } |
| 1057 | |
| 1058 | // |
| 1059 | // Set up Splitter layer |
| 1060 | // |
| 1061 | unsigned int splitterDimSizes[4] = { inputShape[0], inputShape[1], inputShape[2], inputShape[3] }; |
| 1062 | splitterDimSizes[channelsIndex] /= numGroups; // split in depth |
| 1063 | |
| 1064 | TensorInfo splitterOutputInfo(4, |
| 1065 | splitterDimSizes, |
| 1066 | inputInfo.GetDataType(), |
| 1067 | inputInfo.GetQuantizationScale(), |
| 1068 | inputInfo.GetQuantizationOffset()); |
| 1069 | |
| 1070 | std::vector<std::reference_wrapper<TensorInfo>> splitterOutputInfos(numGroups, std::ref(splitterOutputInfo)); |
| 1071 | |
| 1072 | ViewsDescriptor splitterDesc(numGroups); |
| 1073 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1074 | { |
| 1075 | splitterDesc.SetViewOriginCoord(group, channelsIndex, splitterDimSizes[channelsIndex] * group); |
| 1076 | for (unsigned int dimIdx = 0u; dimIdx < 4u; dimIdx++) |
| 1077 | { |
| 1078 | splitterDesc.SetViewSize(group, dimIdx, splitterDimSizes[dimIdx]); |
| 1079 | } |
| 1080 | } |
| 1081 | |
| 1082 | bool isSupported = false; |
| 1083 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1084 | IsSplitterSupported, |
| 1085 | data.m_Backends, |
| 1086 | isSupported, |
| 1087 | inputInfo, |
| 1088 | splitterOutputInfos, |
| 1089 | splitterDesc); |
| 1090 | if (!isSupported) |
| 1091 | { |
| 1092 | return false; |
| 1093 | } |
| 1094 | |
| 1095 | IConnectableLayer* splitterLayer = data.m_Network->AddSplitterLayer(splitterDesc); |
| 1096 | if (!splitterLayer) |
| 1097 | { |
| 1098 | return Fail("%s: Failed to add SplitterLayer", __func__); |
| 1099 | } |
| 1100 | |
| 1101 | input.Connect(splitterLayer->GetInputSlot(0)); |
| 1102 | for (unsigned int group = 0u; group < splitterLayer->GetNumOutputSlots(); ++group) |
| 1103 | { |
| 1104 | splitterLayer->GetOutputSlot(group).SetTensorInfo(splitterOutputInfo); |
| 1105 | } |
| 1106 | |
| 1107 | // |
| 1108 | // Set up Convolution2d layers for each group |
| 1109 | // |
| 1110 | |
| 1111 | // Set up group tensor shapes |
| 1112 | TensorShape groupInputShape(inputShape); |
| 1113 | groupInputShape[channelsIndex] = channelsPerGroup; |
| 1114 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1115 | TensorShape groupWeightsShape(weightsShape); |
| 1116 | groupWeightsShape[0] /= channelMultiplier * numGroups; |
| 1117 | |
| 1118 | TensorShape groupBiasesShape({ 1 }); |
| 1119 | |
| 1120 | // Set up group tensor infos |
| 1121 | TensorInfo groupInputInfo(inputInfo); |
| 1122 | groupInputInfo.SetShape(groupInputShape); |
| 1123 | |
| 1124 | const TensorInfo& weightsInfo = weights.GetInfo(); |
| 1125 | TensorInfo groupWeightsInfo(weightsInfo); |
| 1126 | groupWeightsInfo.SetShape(groupWeightsShape); |
| 1127 | |
| 1128 | const TensorInfo& biasesInfo = biases.GetInfo(); |
| 1129 | TensorInfo groupBiasesInfo(biasesInfo); |
| 1130 | groupBiasesInfo.SetShape(groupBiasesShape); |
| 1131 | |
| 1132 | TensorInfo groupOutputInfo(outputInfo); |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1133 | |
| 1134 | TensorShape groupOutputShape(outputShape); |
| 1135 | const bool isDynamic = IsDynamicTensor(outputInfo); |
| 1136 | if (!isDynamic) |
| 1137 | { |
| 1138 | groupOutputShape[channelsIndex] = 1; |
| 1139 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1140 | groupOutputInfo.SetShape(groupOutputShape); |
| 1141 | |
| 1142 | const unsigned int weightsDataTypeSize = GetDataTypeSize(groupWeightsInfo.GetDataType()); |
| 1143 | const unsigned int biasesDataTypeSize = GetDataTypeSize(groupBiasesInfo.GetDataType()); |
| 1144 | |
| 1145 | std::vector<IConnectableLayer*> convLayers(numGroups * channelMultiplier, nullptr); |
| 1146 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1147 | { |
| 1148 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1149 | { |
| 1150 | auto index = group * channelMultiplier + m; |
| 1151 | |
| 1152 | const unsigned int weightsDataOffset = groupWeightsShape.GetNumElements() * index * weightsDataTypeSize; |
| 1153 | const unsigned int biasesDataOffset = groupBiasesShape.GetNumElements() * index * biasesDataTypeSize; |
| 1154 | |
| 1155 | if (weightsInfo.HasPerAxisQuantization()) |
| 1156 | { |
| 1157 | // Extract per-axis quantization scales for group weights |
| 1158 | const std::vector<float>& weightsQuantScales = weightsInfo.GetQuantizationScales(); |
| 1159 | groupWeightsInfo.SetQuantizationScales( |
| 1160 | std::vector<float>(weightsQuantScales.begin() + index, |
| 1161 | weightsQuantScales.begin() + index + groupWeightsShape[0])); |
| 1162 | |
| 1163 | // Extract per-axis quantization scales for group biases |
| 1164 | const std::vector<float>& biasesQuantScales = biasesInfo.GetQuantizationScales(); |
| 1165 | groupBiasesInfo.SetQuantizationScales( |
| 1166 | std::vector<float>(biasesQuantScales.begin() + index, |
| 1167 | biasesQuantScales.begin() + index + groupWeightsShape[0])); |
| 1168 | } |
| 1169 | |
| 1170 | // Extract weights and biases data for current group convolution |
| 1171 | ConstTensor groupWeights(groupWeightsInfo, |
| 1172 | static_cast<const void *>(reinterpret_cast<const char *>(weights.GetMemoryArea()) + |
| 1173 | weightsDataOffset)); |
| 1174 | ConstTensor groupBiases(groupBiasesInfo, |
| 1175 | static_cast<const void *>(reinterpret_cast<const char *>(biases.GetMemoryArea()) + |
| 1176 | biasesDataOffset)); |
| 1177 | |
| 1178 | isSupported = false; |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1179 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1180 | { |
| 1181 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1182 | IsConvolution2dSupported, |
| 1183 | data.m_Backends, |
| 1184 | isSupported, |
| 1185 | groupInputInfo, |
| 1186 | outputInfo, |
| 1187 | desc, |
| 1188 | groupWeightsInfo, |
| 1189 | Optional<TensorInfo>(groupBiasesInfo)); |
| 1190 | }; |
| 1191 | |
| 1192 | if(!isDynamic) |
| 1193 | { |
| 1194 | validateFunc(groupOutputInfo, isSupported); |
| 1195 | } |
| 1196 | else |
| 1197 | { |
| 1198 | isSupported = AreDynamicTensorsSupported(); |
| 1199 | } |
| 1200 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1201 | if (!isSupported) |
| 1202 | { |
| 1203 | return false; |
| 1204 | } |
| 1205 | |
| 1206 | IConnectableLayer* convLayer = |
| 1207 | data.m_Network->AddConvolution2dLayer(desc, groupWeights, Optional<ConstTensor>(groupBiases)); |
| 1208 | if (!convLayer) |
| 1209 | { |
| 1210 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 1211 | } |
| 1212 | |
| 1213 | splitterLayer->GetOutputSlot(group).Connect(convLayer->GetInputSlot(0)); |
| 1214 | convLayer->GetOutputSlot(0).SetTensorInfo(groupOutputInfo); |
| 1215 | |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1216 | if(isDynamic) |
| 1217 | { |
| 1218 | convLayer->GetOutputSlot(0).IsTensorInfoSet(); |
| 1219 | |
| 1220 | validateFunc(convLayer->GetOutputSlot(0).GetTensorInfo(), isSupported); |
| 1221 | |
| 1222 | outputInfo = convLayer->GetOutputSlot(0).GetTensorInfo(); |
| 1223 | |
| 1224 | if (!isSupported) |
| 1225 | { |
| 1226 | return false; |
| 1227 | } |
| 1228 | } |
| 1229 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1230 | convLayers[index] = convLayer; |
| 1231 | } |
| 1232 | } |
| 1233 | |
| 1234 | // |
| 1235 | // Set up Concat layer |
| 1236 | // |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1237 | ConcatDescriptor concatDescriptor; |
| 1238 | // Equivalent to outputShape[channelsIndex], but we can't know the outputShape in the case of dynamic tensors |
| 1239 | concatDescriptor = ConcatDescriptor(weightsShape[0]); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1240 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1241 | { |
| 1242 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1243 | { |
| 1244 | auto index = group * channelMultiplier + m; |
| 1245 | concatDescriptor.SetViewOriginCoord(index, channelsIndex, index); |
| 1246 | concatDescriptor.SetConcatAxis(channelsIndex); |
| 1247 | } |
| 1248 | } |
| 1249 | |
| 1250 | isSupported = false; |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1251 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1252 | IsConcatSupported, |
| 1253 | data.m_Backends, |
| 1254 | isSupported, |
| 1255 | std::vector<const TensorInfo*>(numGroups * channelMultiplier, &groupOutputInfo), |
| 1256 | outputInfo, |
| 1257 | concatDescriptor); |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1258 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1259 | if (!isSupported) |
| 1260 | { |
| 1261 | return false; |
| 1262 | } |
| 1263 | |
| 1264 | IConnectableLayer* concatLayer = data.m_Network->AddConcatLayer(concatDescriptor); |
| 1265 | if (!concatLayer) |
| 1266 | { |
| 1267 | return Fail("%s: AddConcatLayer failed", __func__); |
| 1268 | } |
| 1269 | |
| 1270 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1271 | { |
| 1272 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1273 | { |
| 1274 | auto index = group * channelMultiplier + m; |
| 1275 | convLayers[index]->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(index)); |
| 1276 | } |
| 1277 | } |
| 1278 | concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1279 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 1280 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *concatLayer, model, |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1281 | data, nullptr, nullptr, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1282 | } |
| 1283 | |
| 1284 | template<typename HalPolicy, |
| 1285 | typename HalOperation = typename HalPolicy::Operation, |
| 1286 | typename HalModel = typename HalPolicy::Model> |
| 1287 | bool ConvertInstanceNormalization(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1288 | { |
| 1289 | using HalOperand = typename HalPolicy::Operand; |
| 1290 | using HalOperandType = typename HalPolicy::OperandType; |
| 1291 | |
| 1292 | ALOGV("HalPolicy::ConvertInstanceNormalization()"); |
| 1293 | |
| 1294 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1295 | if (!input.IsValid()) |
| 1296 | { |
| 1297 | return Fail("%s: Operation has an invalid input 0", __func__); |
| 1298 | } |
| 1299 | |
| 1300 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1301 | if (!output) |
| 1302 | { |
| 1303 | return Fail("%s: Operation has an invalid output", __func__); |
| 1304 | } |
| 1305 | |
| 1306 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1307 | |
| 1308 | // Determine data type of input tensor |
| 1309 | HalOperandType inputType; |
| 1310 | if (!GetOperandType<HalPolicy>(operation, 0, model, inputType)) |
| 1311 | { |
| 1312 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1313 | } |
| 1314 | |
| 1315 | InstanceNormalizationDescriptor desc; |
| 1316 | |
| 1317 | // Read gamma, beta & epsilon |
| 1318 | if (inputType == HalOperandType::TENSOR_FLOAT16) |
| 1319 | { |
| 1320 | Half fp16Gamma; |
| 1321 | Half fp16Beta; |
| 1322 | Half fp16Epsilon; |
| 1323 | |
| 1324 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, fp16Gamma, model, data) || |
| 1325 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::FLOAT16, fp16Beta, model, data) || |
| 1326 | !GetInputScalar<HalPolicy>(operation, 3, HalOperandType::FLOAT16, fp16Epsilon, model, data)) |
| 1327 | { |
| 1328 | return Fail("%s: Operation has invalid inputs (FLOAT16)", __func__); |
| 1329 | } |
| 1330 | |
| 1331 | desc.m_Gamma = static_cast<float>(fp16Gamma); |
| 1332 | desc.m_Beta = static_cast<float>(fp16Beta); |
| 1333 | desc.m_Eps = static_cast<float>(fp16Epsilon); |
| 1334 | } |
| 1335 | else if (inputType == HalOperandType::TENSOR_FLOAT32) |
| 1336 | { |
| 1337 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT32, desc.m_Gamma, model, data) || |
| 1338 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::FLOAT32, desc.m_Beta, model, data) || |
| 1339 | !GetInputScalar<HalPolicy>(operation, 3, HalOperandType::FLOAT32, desc.m_Eps, model, data)) |
| 1340 | { |
| 1341 | return Fail("%s: Operation has invalid inputs (FLOAT32)", __func__); |
| 1342 | } |
| 1343 | } |
| 1344 | else |
| 1345 | { |
| 1346 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1347 | } |
| 1348 | |
| 1349 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 4, model, data); |
| 1350 | |
| 1351 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1352 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1353 | { |
| 1354 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1355 | IsInstanceNormalizationSupported, |
| 1356 | data.m_Backends, |
| 1357 | isSupported, |
| 1358 | input.GetTensorInfo(), |
| 1359 | outputInfo, |
| 1360 | desc); |
| 1361 | }; |
| 1362 | |
| 1363 | if(IsDynamicTensor(outputInfo)) |
| 1364 | { |
| 1365 | isSupported = AreDynamicTensorsSupported(); |
| 1366 | } |
| 1367 | else |
| 1368 | { |
| 1369 | validateFunc(outputInfo, isSupported); |
| 1370 | } |
| 1371 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1372 | if (!isSupported) |
| 1373 | { |
| 1374 | return false; |
| 1375 | } |
| 1376 | |
| 1377 | IConnectableLayer* layer = data.m_Network->AddInstanceNormalizationLayer(desc); |
| 1378 | input.Connect(layer->GetInputSlot(0)); |
| 1379 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1380 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1381 | } |
| 1382 | |
| 1383 | template<typename HalPolicy, |
| 1384 | typename HalOperation = typename HalPolicy::Operation, |
| 1385 | typename HalModel = typename HalPolicy::Model> |
| 1386 | bool ConvertLogSoftmax(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1387 | { |
| 1388 | using HalOperand = typename HalPolicy::Operand; |
| 1389 | using HalOperandType = typename HalPolicy::OperandType; |
| 1390 | |
| 1391 | ALOGV("HalPolicy::ConvertLogSoftmax()"); |
| 1392 | |
| 1393 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1394 | if (!input.IsValid()) |
| 1395 | { |
| 1396 | return Fail("%s: Failed to read input 0", __func__); |
| 1397 | } |
| 1398 | |
| 1399 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1400 | if (!output) |
| 1401 | { |
| 1402 | return Fail("%s: Failed to read output", __func__); |
| 1403 | } |
| 1404 | |
| 1405 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1406 | |
| 1407 | // Determine data type of input tensor |
| 1408 | HalOperandType inputType; |
| 1409 | if (!GetOperandType<HalPolicy>(operation, 0, model, inputType)) |
| 1410 | { |
| 1411 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1412 | } |
| 1413 | |
| 1414 | LogSoftmaxDescriptor descriptor; |
| 1415 | |
| 1416 | // Read beta |
| 1417 | if (inputType == HalOperandType::TENSOR_FLOAT16) |
| 1418 | { |
| 1419 | Half fp16Beta; |
| 1420 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, fp16Beta, model, data)) |
| 1421 | { |
| 1422 | return Fail("%s: Failed to read input 1 (FLOAT16)", __func__); |
| 1423 | } |
| 1424 | |
| 1425 | descriptor.m_Beta = static_cast<float>(fp16Beta); |
| 1426 | } |
| 1427 | else if (inputType == HalOperandType::TENSOR_FLOAT32) |
| 1428 | { |
| 1429 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT32, descriptor.m_Beta, model, data)) |
| 1430 | { |
| 1431 | return Fail("%s: Failed to read input 1 (FLOAT32)", __func__); |
| 1432 | } |
| 1433 | } |
| 1434 | else |
| 1435 | { |
| 1436 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1437 | } |
| 1438 | |
| 1439 | // Read axis |
| 1440 | if (!GetInputInt32<HalPolicy>(operation, 2, descriptor.m_Axis, model, data)) |
| 1441 | { |
| 1442 | return Fail("%s: Failed to read input 2", __func__); |
| 1443 | } |
| 1444 | |
| 1445 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1446 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1447 | { |
| 1448 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1449 | IsLogSoftmaxSupported, |
| 1450 | data.m_Backends, |
| 1451 | isSupported, |
| 1452 | input.GetTensorInfo(), |
| 1453 | outputInfo, |
| 1454 | descriptor); |
| 1455 | }; |
| 1456 | |
| 1457 | if(IsDynamicTensor(outputInfo)) |
| 1458 | { |
| 1459 | isSupported = AreDynamicTensorsSupported(); |
| 1460 | } |
| 1461 | else |
| 1462 | { |
| 1463 | validateFunc(outputInfo, isSupported); |
| 1464 | } |
| 1465 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1466 | if (!isSupported) |
| 1467 | { |
| 1468 | return false; |
| 1469 | } |
| 1470 | |
| 1471 | IConnectableLayer* layer = data.m_Network->AddLogSoftmaxLayer(descriptor); |
| 1472 | if (!layer) |
| 1473 | { |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1474 | return Fail("%s: Could not add the LogSoftmaxLayer", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1475 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1476 | input.Connect(layer->GetInputSlot(0)); |
| 1477 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1478 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1479 | } |
| 1480 | |
| 1481 | template<typename HalPolicy, |
| 1482 | typename HalOperation = typename HalPolicy::Operation, |
| 1483 | typename HalModel = typename HalPolicy::Model> |
| 1484 | bool ConvertMaximum(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1485 | { |
| 1486 | using HalOperand = typename HalPolicy::Operand; |
| 1487 | |
| 1488 | ALOGV("HalPolicy::ConvertMaximum()"); |
| 1489 | |
| 1490 | LayerInputHandle input0 = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1491 | LayerInputHandle input1 = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 1492 | |
| 1493 | if (!input0.IsValid() || !input1.IsValid()) |
| 1494 | { |
| 1495 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1496 | } |
| 1497 | |
| 1498 | const HalOperand* outputOperand = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1499 | if (!outputOperand) |
| 1500 | { |
| 1501 | return Fail("%s: Could not read output", __func__); |
| 1502 | } |
| 1503 | |
| 1504 | const TensorInfo& outInfo = GetTensorInfoForOperand(*outputOperand); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1505 | |
| 1506 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1507 | auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported) |
| 1508 | { |
| 1509 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1510 | IsMaximumSupported, |
| 1511 | data.m_Backends, |
| 1512 | isSupported, |
| 1513 | input0.GetTensorInfo(), |
| 1514 | input1.GetTensorInfo(), |
| 1515 | outInfo); |
| 1516 | }; |
| 1517 | |
| 1518 | if(IsDynamicTensor(outInfo)) |
| 1519 | { |
| 1520 | isSupported = AreDynamicTensorsSupported(); |
| 1521 | } |
| 1522 | else |
| 1523 | { |
| 1524 | validateFunc(outInfo, isSupported); |
| 1525 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1526 | |
| 1527 | if (!isSupported) |
| 1528 | { |
| 1529 | return false; |
| 1530 | } |
| 1531 | |
| 1532 | IConnectableLayer* layer = data.m_Network->AddMaximumLayer(); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1533 | if (!layer) |
| 1534 | { |
| 1535 | return Fail("%s: Could not add the MaximumLayer", __func__); |
| 1536 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1537 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 1538 | if (!isReshapeSupported) |
| 1539 | { |
| 1540 | return false; |
| 1541 | } |
| 1542 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1543 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1544 | } |
| 1545 | |
| 1546 | template<typename HalPolicy, |
| 1547 | typename HalOperation = typename HalPolicy::Operation, |
| 1548 | typename HalModel = typename HalPolicy::Model> |
| 1549 | bool ConvertMinimum(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1550 | { |
| 1551 | using HalOperand = typename HalPolicy::Operand; |
| 1552 | |
| 1553 | ALOGV("HalPolicy::ConvertMinimum()"); |
| 1554 | |
| 1555 | LayerInputHandle input0 = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1556 | LayerInputHandle input1 = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 1557 | |
| 1558 | if (!input0.IsValid() || !input1.IsValid()) |
| 1559 | { |
| 1560 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1561 | } |
| 1562 | |
| 1563 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1564 | if (!output) |
| 1565 | { |
| 1566 | return Fail("%s: Could not read output 0", __func__); |
| 1567 | } |
| 1568 | |
| 1569 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1570 | |
| 1571 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1572 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1573 | { |
| 1574 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1575 | IsMinimumSupported, |
| 1576 | data.m_Backends, |
| 1577 | isSupported, |
| 1578 | input0.GetTensorInfo(), |
| 1579 | input1.GetTensorInfo(), |
| 1580 | outputInfo); |
| 1581 | }; |
| 1582 | |
| 1583 | if(IsDynamicTensor(outputInfo)) |
| 1584 | { |
| 1585 | isSupported = AreDynamicTensorsSupported(); |
| 1586 | } |
| 1587 | else |
| 1588 | { |
| 1589 | validateFunc(outputInfo, isSupported); |
| 1590 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1591 | |
| 1592 | if (!isSupported) |
| 1593 | { |
| 1594 | return false; |
| 1595 | } |
| 1596 | |
| 1597 | IConnectableLayer* const layer = data.m_Network->AddMinimumLayer(); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1598 | if (!layer) |
| 1599 | { |
| 1600 | return Fail("%s: Could not add the MinimumLayer", __func__); |
| 1601 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1602 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 1603 | if (!isReshapeSupported) |
| 1604 | { |
| 1605 | return false; |
| 1606 | } |
| 1607 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1608 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1609 | } |
| 1610 | |
| 1611 | template<typename HalPolicy, |
| 1612 | typename HalOperation = typename HalPolicy::Operation, |
| 1613 | typename HalModel = typename HalPolicy::Model> |
| 1614 | bool ConvertPadV2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1615 | { |
| 1616 | using HalOperand = typename HalPolicy::Operand; |
| 1617 | using HalOperandType = typename HalPolicy::OperandType; |
| 1618 | |
| 1619 | ALOGV("HalPolicy::ConvertPadV2()"); |
| 1620 | |
| 1621 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1622 | if (!input.IsValid()) |
| 1623 | { |
| 1624 | return Fail("%s: Could not read input 0", __func__); |
| 1625 | } |
| 1626 | |
| 1627 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1628 | if (!output) |
| 1629 | { |
| 1630 | return Fail("%s: Could not read output", __func__); |
| 1631 | } |
| 1632 | |
| 1633 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1634 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 1635 | |
| 1636 | PadDescriptor descriptor; |
| 1637 | if (!ConvertPaddings<HalPolicy>(operation, model, data, rank, descriptor)) |
| 1638 | { |
| 1639 | return Fail("%s: Could not convert paddings", __func__); |
| 1640 | } |
| 1641 | |
| 1642 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1643 | |
| 1644 | // Determine type of padding value |
| 1645 | HalOperandType operandType0; |
| 1646 | HalOperandType operandType2; |
| 1647 | |
| 1648 | if (!GetOperandType<HalPolicy>(operation, 0, model, operandType0) || |
| 1649 | !GetOperandType<HalPolicy>(operation, 2, model, operandType2)) |
| 1650 | { |
| 1651 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1652 | } |
| 1653 | |
| 1654 | // Read value to use for padding |
| 1655 | if (operandType0 == HalOperandType::TENSOR_FLOAT16 && operandType2 == HalOperandType::FLOAT16) |
| 1656 | { |
| 1657 | Half f16PadValue; |
| 1658 | if (!GetInputScalar<HalPolicy>(operation, 2, operandType2, f16PadValue, model, data)) |
| 1659 | { |
| 1660 | return Fail("%s: Could not read input 2 (FLOAT16)", __func__); |
| 1661 | } |
| 1662 | |
| 1663 | descriptor.m_PadValue = f16PadValue; |
| 1664 | } |
| 1665 | else if (operandType0 == HalOperandType::TENSOR_FLOAT32 && operandType2 == HalOperandType::FLOAT32) |
| 1666 | { |
| 1667 | if (!GetInputFloat32<HalPolicy>(operation, 2, descriptor.m_PadValue, model, data)) |
| 1668 | { |
| 1669 | return Fail("%s: Could not read input 2 (FLOAT32)", __func__); |
| 1670 | } |
| 1671 | } |
Teresa Charlin | d3381d5 | 2021-06-02 18:35:16 +0100 | [diff] [blame] | 1672 | else if (isQuantizedOperand(operandType0) && operandType2 == HalOperandType::INT32) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1673 | { |
| 1674 | int32_t intPadValue = 0; |
| 1675 | if (!GetInputInt32<HalPolicy>(operation, 2, intPadValue, model, data)) |
| 1676 | { |
| 1677 | return Fail("%s: Could not read input 2 (INT32)", __func__); |
| 1678 | } |
| 1679 | descriptor.m_PadValue = intPadValue; |
| 1680 | } |
| 1681 | else |
| 1682 | { |
| 1683 | return Fail("%s: Operation has invalid inputs: type mismatch", __func__); |
| 1684 | } |
| 1685 | |
| 1686 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1687 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1688 | { |
| 1689 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1690 | IsPadSupported, |
| 1691 | data.m_Backends, |
| 1692 | isSupported, |
| 1693 | inputInfo, |
| 1694 | outputInfo, |
| 1695 | descriptor); |
| 1696 | }; |
| 1697 | |
| 1698 | if(IsDynamicTensor(outputInfo)) |
| 1699 | { |
| 1700 | isSupported = AreDynamicTensorsSupported(); |
| 1701 | } |
| 1702 | else |
| 1703 | { |
| 1704 | validateFunc(outputInfo, isSupported); |
| 1705 | } |
| 1706 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1707 | if (!isSupported) |
| 1708 | { |
| 1709 | return false; |
| 1710 | } |
| 1711 | |
| 1712 | IConnectableLayer* const layer = data.m_Network->AddPadLayer(descriptor); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1713 | if (!layer) |
| 1714 | { |
| 1715 | return Fail("%s: Could not add the PadLayer", __func__); |
| 1716 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1717 | input.Connect(layer->GetInputSlot(0)); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1718 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1719 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1720 | } |
| 1721 | |
| 1722 | template<typename HalPolicy, |
| 1723 | typename HalOperation = typename HalPolicy::Operation, |
| 1724 | typename HalModel = typename HalPolicy::Model> |
| 1725 | bool ConvertPrelu(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1726 | { |
| 1727 | using HalOperand = typename HalPolicy::Operand; |
| 1728 | |
| 1729 | ALOGV("HalPolicy::ConvertPrelu()"); |
| 1730 | |
| 1731 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1732 | LayerInputHandle alpha = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 1733 | |
| 1734 | if (!input.IsValid() || !alpha.IsValid()) |
| 1735 | { |
| 1736 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1737 | } |
| 1738 | |
| 1739 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1740 | |
| 1741 | if (!output) |
| 1742 | { |
| 1743 | return Fail("%s: Could not read output", __func__); |
| 1744 | } |
| 1745 | |
| 1746 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1747 | const TensorInfo& alphaInfo = alpha.GetTensorInfo(); |
| 1748 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1749 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1750 | bool isSupported = false; |
| 1751 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1752 | { |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1753 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1754 | IsPreluSupported, |
| 1755 | data.m_Backends, |
| 1756 | isSupported, |
| 1757 | inputInfo, |
| 1758 | alphaInfo, |
| 1759 | outputInfo); |
| 1760 | }; |
| 1761 | |
| 1762 | if(IsDynamicTensor(outputInfo)) |
| 1763 | { |
| 1764 | isSupported = AreDynamicTensorsSupported(); |
| 1765 | } |
| 1766 | else |
| 1767 | { |
| 1768 | validateFunc(outputInfo, isSupported); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1769 | } |
| 1770 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1771 | if (!isSupported) |
| 1772 | { |
| 1773 | return false; |
| 1774 | } |
| 1775 | |
| 1776 | IConnectableLayer* const layer = data.m_Network->AddPreluLayer(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1777 | if (!layer) |
| 1778 | { |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1779 | return Fail("%s: Could not add the PreluLayer", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1780 | } |
| 1781 | |
| 1782 | bool isReshapeSupported = BroadcastTensor(input, alpha, layer, data); |
| 1783 | if (!isReshapeSupported) |
| 1784 | { |
| 1785 | return false; |
| 1786 | } |
| 1787 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1788 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1789 | } |
| 1790 | |
| 1791 | template<typename HalPolicy, |
| 1792 | typename HalOperation = typename HalPolicy::Operation, |
| 1793 | typename HalModel = typename HalPolicy::Model> |
| 1794 | bool ConvertQuantize(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1795 | { |
| 1796 | using HalOperand = typename HalPolicy::Operand; |
| 1797 | |
| 1798 | ALOGV("HalPolicy::ConvertQuantize()"); |
| 1799 | |
| 1800 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1801 | if (!input.IsValid()) |
| 1802 | { |
| 1803 | return Fail("%s: Operation has invalid input", __func__); |
| 1804 | } |
| 1805 | |
| 1806 | const HalOperand* const outputOperand = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1807 | if (!outputOperand) |
| 1808 | { |
| 1809 | return Fail("%s: Operation has invalid outputs", __func__); |
| 1810 | } |
| 1811 | |
| 1812 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1813 | |
| 1814 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1815 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1816 | { |
| 1817 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1818 | IsQuantizeSupported, |
| 1819 | data.m_Backends, |
| 1820 | isSupported, |
| 1821 | input.GetTensorInfo(), |
| 1822 | outputInfo); |
| 1823 | }; |
| 1824 | |
| 1825 | if(IsDynamicTensor(outputInfo)) |
| 1826 | { |
| 1827 | isSupported = AreDynamicTensorsSupported(); |
| 1828 | } |
| 1829 | else |
| 1830 | { |
| 1831 | validateFunc(outputInfo, isSupported); |
| 1832 | } |
| 1833 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1834 | if (!isSupported) |
| 1835 | { |
| 1836 | return false; |
| 1837 | } |
| 1838 | |
| 1839 | IConnectableLayer* const layer = data.m_Network->AddQuantizeLayer(); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1840 | if (!layer) |
| 1841 | { |
| 1842 | return Fail("%s: Could not add the QuantizeLayer", __func__); |
| 1843 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1844 | input.Connect(layer->GetInputSlot(0)); |
| 1845 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1846 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1847 | } |
| 1848 | |
| 1849 | template<typename HalPolicy, |
| 1850 | typename HalOperation = typename HalPolicy::Operation, |
| 1851 | typename HalModel = typename HalPolicy::Model> |
Sadik Armagan | 813f230 | 2020-05-19 14:10:30 +0100 | [diff] [blame] | 1852 | bool ConvertQuantized16BitLstm(const HalOperation& operation, const HalModel& model, ConversionData& data) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1853 | { |
| 1854 | using HalOperand = typename HalPolicy::Operand; |
| 1855 | |
Sadik Armagan | 813f230 | 2020-05-19 14:10:30 +0100 | [diff] [blame] | 1856 | ALOGV("HalPolicy::ConvertQuantized16BitLstm()"); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1857 | |
| 1858 | //Inputs: |
| 1859 | // 0: The input: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBatches, inputSize] |
| 1860 | // specifying the input to the LSTM cell. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 1861 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1862 | if (!input.IsValid()) |
| 1863 | { |
| 1864 | return Fail("%s: Could not read input 0: input", __func__); |
| 1865 | } |
| 1866 | |
| 1867 | //13: The previous cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape |
| 1868 | // [numBatches, outputSize] specifying the cell state from the previous time step of the LSTM cell. |
| 1869 | // It is quantized using a quantization range of -2^4, 2^4 * 32767/32768. |
| 1870 | LayerInputHandle previousCellStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 13, model, data); |
| 1871 | if (!previousCellStateIn.IsValid()) |
| 1872 | { |
| 1873 | return Fail("%s: Could not read input 13: previousCellStateIn", __func__); |
| 1874 | } |
| 1875 | |
| 1876 | // 14: The previous output state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1877 | // [numBathes, outputSize] specifying the output of the LSTM cell from previous time-step. Tensor |
| 1878 | // is quantized with a fixed quantization range of -1, 127/128. |
| 1879 | LayerInputHandle previousOutputIn = ConvertToLayerInputHandle<HalPolicy>(operation, 14, model, data); |
| 1880 | if (!previousOutputIn.IsValid()) |
| 1881 | { |
| 1882 | return Fail("%s: Could not read input 14: previousOutputIn", __func__); |
| 1883 | } |
| 1884 | |
| 1885 | // Get the input tensors: |
| 1886 | // 1: The input-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1887 | // [outputSize, inputSize] specifying input-to-input part of weights for fully-connected layer inside the |
| 1888 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1889 | const ConstTensorPin inputToInputWeightsPin = |
| 1890 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 1891 | |
| 1892 | // 2: The input-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1893 | // [outputSize, inputSize] specifying input-to-forget part of weights for fully-connected layer inside the |
| 1894 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1895 | const ConstTensorPin inputToForgetWeightsPin = |
| 1896 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 1897 | |
| 1898 | // 3: The input-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1899 | // [outputSize, inputSize] specifying input-to-cell part of weights for fully-connected layer inside the |
| 1900 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1901 | const ConstTensorPin inputToCellWeightsPin = |
| 1902 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 3, model, data); |
| 1903 | |
| 1904 | // 4: The input-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1905 | // [outputSize, inputSize] specifying input-to-output part of weights for fully-connected layer inside the |
| 1906 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1907 | const ConstTensorPin inputToOutputWeightsPin = |
| 1908 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 4, model, data); |
| 1909 | |
| 1910 | // 5: The recurrent-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1911 | // [outputSize, outputSize] specifying recurrent-to-input part of weights for fully-connected layer inside |
| 1912 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1913 | const ConstTensorPin recurrentToInputWeightsPin = |
| 1914 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 5, model, data); |
| 1915 | |
| 1916 | // 6: The recurrent-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1917 | // [outputSize, outputSize] specifying recurrent-to-forget part of weights for fully-connected layer inside |
| 1918 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1919 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 1920 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 6, model, data); |
| 1921 | |
| 1922 | // 7: The recurrent-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1923 | // [outputSize, outputSize] specifying recurrent-to-cell part of weights for fully-connected layer inside |
| 1924 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1925 | const ConstTensorPin recurrentToCellWeightsPin = |
| 1926 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 7, model, data); |
| 1927 | |
| 1928 | // 8: The recurrent-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1929 | // [outputSize, outputSize] specifying recurrent-to-output part of weights for fully-connected layer inside |
| 1930 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1931 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 1932 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 8, model, data); |
| 1933 | |
| 1934 | // 9: The input gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the |
| 1935 | // bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1936 | // of input and weights scales and zeroPoint equal to 0. |
| 1937 | const ConstTensorPin inputGateBiasPin = |
| 1938 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 9, model, data); |
| 1939 | |
| 1940 | // 10: The forget gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 1941 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1942 | // of input and weights scales and zeroPoint equal to 0. |
| 1943 | const ConstTensorPin forgetGateBiasPin = |
| 1944 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 10, model, data); |
| 1945 | |
| 1946 | // 11:The cell bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the bias |
| 1947 | // for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product of input |
| 1948 | // and weights scales and zeroPoint equal to 0. |
| 1949 | const ConstTensorPin cellBiasPin = |
| 1950 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 11, model, data); |
| 1951 | |
| 1952 | // 12:The output gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 1953 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1954 | // of input and weights scales and zeroPoint equal to 0. |
| 1955 | const ConstTensorPin outputGateBiasPin = |
| 1956 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 12, model, data); |
| 1957 | |
| 1958 | if (!inputToInputWeightsPin.IsValid() || |
| 1959 | !inputToForgetWeightsPin.IsValid() || |
| 1960 | !inputToCellWeightsPin.IsValid() || |
| 1961 | !inputToOutputWeightsPin.IsValid() || |
| 1962 | !recurrentToInputWeightsPin.IsValid() || |
| 1963 | !recurrentToForgetWeightsPin.IsValid() || |
| 1964 | !recurrentToCellWeightsPin.IsValid() || |
| 1965 | !recurrentToOutputWeightsPin.IsValid() || |
| 1966 | !inputGateBiasPin.IsValid() || |
| 1967 | !forgetGateBiasPin.IsValid() || |
| 1968 | !cellBiasPin.IsValid() || |
| 1969 | !outputGateBiasPin.IsValid()) |
| 1970 | { |
| 1971 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 1972 | } |
| 1973 | |
| 1974 | // Outputs: |
| 1975 | // 0: The cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape [numBatches, outputSize] |
| 1976 | // which contains a cell state from the current time step. Tensor is quantized using a quantization range |
| 1977 | // of -2^4, 2^4 * 32767/32768. |
| 1978 | const HalOperand* cellStateOut = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1979 | if (!cellStateOut) |
| 1980 | { |
| 1981 | return Fail("%s: Could not read output 0: cellStateOut", __func__); |
| 1982 | } |
| 1983 | |
| 1984 | // 1: The output: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBathes, outputSize] which |
| 1985 | // contains the output value. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 1986 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 1, model); |
| 1987 | if (!output) |
| 1988 | { |
| 1989 | return Fail("%s: Could not read output 1: output", __func__); |
| 1990 | } |
| 1991 | |
| 1992 | // Inputs |
| 1993 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1994 | const TensorInfo& previousCellStateInInfo = previousCellStateIn.GetTensorInfo(); |
| 1995 | const TensorInfo& previousOutputInInfo = previousOutputIn.GetTensorInfo(); |
| 1996 | |
| 1997 | // Outputs |
| 1998 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 1999 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2000 | |
| 2001 | // Dynamic tensors currently not supported |
| 2002 | if (IsDynamicTensor(cellStateOutInfo) || IsDynamicTensor(outputInfo)) |
| 2003 | { |
| 2004 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 2005 | } |
| 2006 | |
| 2007 | QuantizedLstmInputParams params; |
| 2008 | |
| 2009 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 2010 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 2011 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 2012 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 2013 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 2014 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 2015 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 2016 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 2017 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 2018 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 2019 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 2020 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 2021 | |
| 2022 | QuantizedLstmInputParamsInfo paramsInfo; |
| 2023 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 2024 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 2025 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 2026 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 2027 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 2028 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 2029 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 2030 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 2031 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 2032 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 2033 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 2034 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 2035 | |
| 2036 | bool isSupported = false; |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2037 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2038 | { |
| 2039 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2040 | IsQuantizedLstmSupported, |
| 2041 | data.m_Backends, |
| 2042 | isSupported, |
| 2043 | inputInfo, |
| 2044 | previousCellStateInInfo, |
| 2045 | previousOutputInInfo, |
| 2046 | cellStateOutInfo, |
| 2047 | outputInfo, |
| 2048 | paramsInfo); |
| 2049 | }; |
| 2050 | |
| 2051 | bool isDynamic = false; |
| 2052 | if (!IsDynamicTensor(cellStateOutInfo) && |
| 2053 | !IsDynamicTensor(outputInfo)) |
| 2054 | { |
| 2055 | validateFunc(outputInfo, isSupported); |
| 2056 | } |
| 2057 | else |
| 2058 | { |
| 2059 | isDynamic = true; |
| 2060 | isSupported = AreDynamicTensorsSupported(); |
| 2061 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2062 | |
| 2063 | if (!isSupported) |
| 2064 | { |
| 2065 | return false; |
| 2066 | } |
| 2067 | |
| 2068 | IConnectableLayer* const layer = data.m_Network->AddQuantizedLstmLayer(params, "QuantizedLstm"); |
| 2069 | input.Connect(layer->GetInputSlot(0)); |
| 2070 | previousCellStateIn.Connect(layer->GetInputSlot(1)); |
| 2071 | previousOutputIn.Connect(layer->GetInputSlot(2)); |
| 2072 | |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2073 | if (!isDynamic) |
| 2074 | { |
| 2075 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 2076 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data)); |
| 2077 | } |
| 2078 | else |
| 2079 | { |
| 2080 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 2081 | SetupAndTrackLayerOutputSlot<HalPolicy>( |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 2082 | operation, 1, *layer, 1, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2083 | } |
| 2084 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2085 | } |
| 2086 | |
| 2087 | template<typename HalPolicy, |
| 2088 | typename HalOperation = typename HalPolicy::Operation, |
| 2089 | typename HalModel = typename HalPolicy::Model> |
Teresa Charlin | 89cbb3a | 2021-02-11 21:00:47 +0000 | [diff] [blame] | 2090 | bool ConvertReduce(const HalOperation& operation, |
| 2091 | const HalModel& model, |
| 2092 | ConversionData& data, |
| 2093 | ReduceOperation reduceOperation) |
| 2094 | { |
| 2095 | using HalOperand = typename HalPolicy::Operand; |
| 2096 | using HalOperandType = typename HalPolicy::OperandType; |
| 2097 | |
| 2098 | armnn::ReduceDescriptor descriptor; |
| 2099 | descriptor.m_ReduceOperation = reduceOperation; |
| 2100 | |
| 2101 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2102 | if (!input.IsValid()) |
| 2103 | { |
| 2104 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2105 | } |
| 2106 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2107 | |
| 2108 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2109 | if (!output) |
| 2110 | { |
| 2111 | return Fail("%s: Could not read output 0", __func__); |
| 2112 | } |
| 2113 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2114 | |
| 2115 | const HalOperand* axisOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
| 2116 | if (!axisOperand) |
| 2117 | { |
| 2118 | return Fail("%s: Could not read input 1", __func__); |
| 2119 | } |
| 2120 | std::vector<int32_t> axis; |
| 2121 | if (!GetTensorInt32Values<HalPolicy>(*axisOperand, axis, model, data)) |
| 2122 | { |
| 2123 | return Fail("%s: Input 1 has invalid values", __func__); |
| 2124 | } |
| 2125 | |
| 2126 | // Convert the axis to unsigned int and remove duplicates. |
| 2127 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 2128 | std::set<unsigned int> uniqueAxis; |
| 2129 | std::transform(axis.begin(), axis.end(), |
| 2130 | std::inserter(uniqueAxis, uniqueAxis.begin()), |
| 2131 | [rank](int i) -> unsigned int { return (i + rank) % rank; }); |
| 2132 | descriptor.m_vAxis.assign(uniqueAxis.begin(), uniqueAxis.end()); |
| 2133 | |
| 2134 | // Get the "keep dims" flag. |
| 2135 | if (!GetInputScalar<HalPolicy>(operation, 2, HalOperandType::BOOL, descriptor.m_KeepDims, model, data)) |
| 2136 | { |
| 2137 | return Fail("%s: Could not read input 2", __func__); |
| 2138 | } |
| 2139 | |
| 2140 | bool isSupported = false; |
| 2141 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2142 | { |
| 2143 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2144 | IsReduceSupported, |
| 2145 | data.m_Backends, |
| 2146 | isSupported, |
| 2147 | inputInfo, |
| 2148 | outputInfo, |
| 2149 | descriptor); |
| 2150 | }; |
| 2151 | |
| 2152 | if(!IsDynamicTensor(outputInfo)) |
| 2153 | { |
| 2154 | validateFunc(outputInfo, isSupported); |
| 2155 | } |
| 2156 | else |
| 2157 | { |
| 2158 | isSupported = AreDynamicTensorsSupported(); |
| 2159 | } |
| 2160 | |
| 2161 | if (!isSupported) |
| 2162 | { |
| 2163 | return false; |
| 2164 | } |
| 2165 | |
| 2166 | armnn::IConnectableLayer* const layer = data.m_Network->AddReduceLayer(descriptor); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2167 | if (!layer) |
| 2168 | { |
| 2169 | return Fail("%s: Could not add the ReduceLayer", __func__); |
| 2170 | } |
Teresa Charlin | 89cbb3a | 2021-02-11 21:00:47 +0000 | [diff] [blame] | 2171 | input.Connect(layer->GetInputSlot(0)); |
| 2172 | |
| 2173 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2174 | } |
| 2175 | |
| 2176 | template<typename HalPolicy, |
| 2177 | typename HalOperation = typename HalPolicy::Operation, |
| 2178 | typename HalModel = typename HalPolicy::Model> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2179 | bool ConvertResize(const HalOperation& operation, |
| 2180 | const HalModel& model, |
| 2181 | ConversionData& data, |
| 2182 | ResizeMethod resizeMethod) |
| 2183 | { |
| 2184 | using HalOperand = typename HalPolicy::Operand; |
| 2185 | using HalOperandType = typename HalPolicy::OperandType; |
| 2186 | ALOGV("HalPolicy::ConvertResize()"); |
| 2187 | ALOGV("resizeMethod = %s", GetResizeMethodAsCString(resizeMethod)); |
| 2188 | |
| 2189 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2190 | if (!input.IsValid()) |
| 2191 | { |
| 2192 | return Fail("%s: Could not read input 0", __func__); |
| 2193 | } |
| 2194 | |
| 2195 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2196 | if (!output) |
| 2197 | { |
| 2198 | return Fail("%s: Could not read output 0", __func__); |
| 2199 | } |
| 2200 | |
| 2201 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2202 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2203 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2204 | ResizeDescriptor descriptor; |
| 2205 | descriptor.m_Method = resizeMethod; |
| 2206 | descriptor.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 3, model, data); |
| 2207 | |
| 2208 | HalOperandType operandType1; |
| 2209 | HalOperandType operandType2; |
| 2210 | |
| 2211 | if (!GetOperandType<HalPolicy>(operation, 1, model, operandType1) || |
| 2212 | !GetOperandType<HalPolicy>(operation, 2, model, operandType2)) |
| 2213 | { |
| 2214 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2215 | } |
| 2216 | |
| 2217 | if (operandType1 != operandType2) |
| 2218 | { |
| 2219 | return Fail("%s: Operation has invalid inputs. Type of input 1 and 2 should be the same", __func__); |
| 2220 | } |
| 2221 | |
| 2222 | if (operandType1 == HalOperandType::INT32) |
| 2223 | { |
| 2224 | // Case 1: resizing by shape |
| 2225 | int32_t targetWidth = 0; |
| 2226 | int32_t targetHeight = 0; |
| 2227 | |
| 2228 | if (!GetInputInt32<HalPolicy>(operation, 1, targetWidth, model, data) || |
| 2229 | !GetInputInt32<HalPolicy>(operation, 2, targetHeight, model, data)) |
| 2230 | { |
| 2231 | return Fail("%s: Operation has invalid inputs for resizing by shape", __func__); |
| 2232 | } |
| 2233 | |
| 2234 | if (targetWidth < 0 || targetHeight < 0) |
| 2235 | { |
| 2236 | return Fail("%s: Operation has invalid inputs for resizing by shape. " |
| 2237 | "Target width/height cannot be < 0", __func__); |
| 2238 | } |
| 2239 | |
| 2240 | descriptor.m_TargetWidth = static_cast<uint32_t>(targetWidth); |
| 2241 | descriptor.m_TargetHeight = static_cast<uint32_t>(targetHeight); |
| 2242 | } |
| 2243 | else if (operandType1 == HalOperandType::FLOAT32) |
| 2244 | { |
| 2245 | // Case 2: resizing by scale |
| 2246 | float widthScale = 1.0f; |
| 2247 | float heightScale = 1.0f; |
| 2248 | |
| 2249 | if (!GetInputFloat32<HalPolicy>(operation, 1, widthScale, model, data) || |
| 2250 | !GetInputFloat32<HalPolicy>(operation, 2, heightScale, model, data)) |
| 2251 | { |
| 2252 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 2253 | } |
| 2254 | |
| 2255 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 2256 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 2257 | |
| 2258 | float width = inputShape[dataLayoutIndexed.GetWidthIndex()]; |
| 2259 | float height = inputShape[dataLayoutIndexed.GetHeightIndex()]; |
| 2260 | |
| 2261 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 2262 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 2263 | } |
| 2264 | else if (operandType1 == HalOperandType::FLOAT16) |
| 2265 | { |
| 2266 | Half widthScale; |
| 2267 | Half heightScale; |
| 2268 | |
| 2269 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, widthScale, model, data) || |
| 2270 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::FLOAT16, heightScale, model, data)) |
| 2271 | { |
| 2272 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 2273 | } |
| 2274 | |
| 2275 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 2276 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 2277 | |
| 2278 | Half width = static_cast<Half>(inputShape[dataLayoutIndexed.GetWidthIndex()]); |
| 2279 | Half height = static_cast<Half>(inputShape[dataLayoutIndexed.GetHeightIndex()]); |
| 2280 | |
| 2281 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 2282 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 2283 | } |
| 2284 | else |
| 2285 | { |
| 2286 | return Fail("%s: Operand has invalid data type for resizing by scale", __func__); |
| 2287 | } |
| 2288 | |
David Monahan | f057e6f | 2020-06-22 09:55:23 +0100 | [diff] [blame] | 2289 | descriptor.m_AlignCorners = GetOptionalBool<HalPolicy>(operation, 4, model, data); |
| 2290 | descriptor.m_HalfPixelCenters = GetOptionalBool<HalPolicy>(operation, 5, model, data); |
David Monahan | 51e0b13 | 2020-04-20 16:12:06 +0100 | [diff] [blame] | 2291 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2292 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2293 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2294 | { |
| 2295 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2296 | IsResizeSupported, |
| 2297 | data.m_Backends, |
| 2298 | isSupported, |
| 2299 | inputInfo, |
| 2300 | outputInfo, |
| 2301 | descriptor); |
| 2302 | }; |
| 2303 | |
| 2304 | if(IsDynamicTensor(outputInfo)) |
| 2305 | { |
| 2306 | isSupported = AreDynamicTensorsSupported(); |
| 2307 | } |
| 2308 | else |
| 2309 | { |
| 2310 | validateFunc(outputInfo, isSupported); |
| 2311 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2312 | |
| 2313 | if (!isSupported) |
| 2314 | { |
| 2315 | return false; |
| 2316 | } |
| 2317 | |
| 2318 | IConnectableLayer* layer = data.m_Network->AddResizeLayer(descriptor); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2319 | if (!layer) |
| 2320 | { |
| 2321 | return Fail("%s: Could not add the ResizeLayer", __func__); |
| 2322 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2323 | input.Connect(layer->GetInputSlot(0)); |
| 2324 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2325 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2326 | } |
| 2327 | |
| 2328 | template<typename HalPolicy, |
| 2329 | typename HalOperation = typename HalPolicy::Operation, |
| 2330 | typename HalModel = typename HalPolicy::Model> |
| 2331 | bool ConvertSpaceToDepth(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2332 | { |
| 2333 | using HalOperand = typename HalPolicy::Operand; |
| 2334 | using HalOperandType = typename HalPolicy::OperandType; |
| 2335 | |
| 2336 | ALOGV("HalPolicy::ConvertSpaceToDepth()"); |
| 2337 | |
| 2338 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2339 | if (!input.IsValid() ) |
| 2340 | { |
| 2341 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2342 | } |
| 2343 | |
| 2344 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2345 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 2346 | if (rank != 4) |
| 2347 | { |
| 2348 | return Fail("%s: Only inputs with rank 4 are supported", __func__); |
| 2349 | } |
| 2350 | |
| 2351 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2352 | if (!output) |
| 2353 | { |
| 2354 | return Fail("%s: Could not read output 0", __func__); |
| 2355 | } |
| 2356 | |
| 2357 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2358 | |
| 2359 | SpaceToDepthDescriptor desc; |
| 2360 | |
| 2361 | GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, desc.m_BlockSize, model, data); |
| 2362 | |
| 2363 | if (desc.m_BlockSize <= 1) |
| 2364 | { |
| 2365 | return Fail("%s: Block size must be at least 1 in all dimensions"); |
| 2366 | } |
| 2367 | |
| 2368 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 2, model, data); |
| 2369 | |
| 2370 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2371 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2372 | { |
| 2373 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2374 | IsSpaceToDepthSupported, |
| 2375 | data.m_Backends, |
| 2376 | isSupported, |
| 2377 | inputInfo, |
| 2378 | outputInfo, |
| 2379 | desc); |
| 2380 | }; |
| 2381 | |
| 2382 | if(IsDynamicTensor(outputInfo)) |
| 2383 | { |
| 2384 | isSupported = AreDynamicTensorsSupported(); |
| 2385 | } |
| 2386 | else |
| 2387 | { |
| 2388 | validateFunc(outputInfo, isSupported); |
| 2389 | } |
| 2390 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2391 | if (!isSupported) |
| 2392 | { |
| 2393 | return false; |
| 2394 | } |
| 2395 | |
| 2396 | IConnectableLayer* const layer = data.m_Network->AddSpaceToDepthLayer(desc); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2397 | if (!layer) |
| 2398 | { |
Mike Kelly | 1b46d13 | 2021-11-03 11:12:45 +0000 | [diff] [blame] | 2399 | return Fail("%s: Could not add the SpaceToDepthLayer", __func__); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2400 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2401 | input.Connect(layer->GetInputSlot(0)); |
| 2402 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2403 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2404 | } |
| 2405 | |
| 2406 | template<typename HalPolicy, |
| 2407 | typename HalOperation = typename HalPolicy::Operation, |
| 2408 | typename HalModel = typename HalPolicy::Model> |
| 2409 | bool ConvertSoftmax(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2410 | { |
| 2411 | using HalOperand = typename HalPolicy::Operand; |
| 2412 | using HalOperandType = typename HalPolicy::OperandType; |
| 2413 | |
| 2414 | ALOGV("HalPolicy::ConvertSoftmax()"); |
| 2415 | |
| 2416 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2417 | if (!input.IsValid()) |
| 2418 | { |
| 2419 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2420 | } |
| 2421 | |
| 2422 | const HalOperand* outputOperand = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2423 | if (!outputOperand) |
| 2424 | { |
| 2425 | return Fail("%s: Operation has no outputs", __func__); |
| 2426 | } |
| 2427 | |
| 2428 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2429 | |
| 2430 | SoftmaxDescriptor desc; |
Teresa Charlin | bf866e2 | 2020-08-09 23:55:01 +0100 | [diff] [blame] | 2431 | HalOperandType outputType = outputOperand->type; |
| 2432 | |
| 2433 | // Read beta value |
| 2434 | if (outputType == HalOperandType::TENSOR_FLOAT16) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2435 | { |
Teresa Charlin | bf866e2 | 2020-08-09 23:55:01 +0100 | [diff] [blame] | 2436 | Half value; |
| 2437 | |
| 2438 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, value, model, data)) |
| 2439 | { |
| 2440 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 2441 | } |
| 2442 | |
| 2443 | desc.m_Beta = static_cast<float>(value); |
| 2444 | } |
| 2445 | else |
| 2446 | { |
| 2447 | if (!GetInputFloat32<HalPolicy>(operation, 1, desc.m_Beta, model, data)) |
| 2448 | { |
| 2449 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 2450 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2451 | } |
| 2452 | |
| 2453 | if (operation.inputs.size() > 2 && !GetInputScalar<HalPolicy>(operation, |
Teresa Charlin | bf866e2 | 2020-08-09 23:55:01 +0100 | [diff] [blame] | 2454 | 2, |
| 2455 | HalOperandType::INT32, |
| 2456 | desc.m_Axis, |
| 2457 | model, |
| 2458 | data)) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2459 | { |
| 2460 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2461 | } |
| 2462 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2463 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2464 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2465 | { |
| 2466 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2467 | IsSoftmaxSupported, |
| 2468 | data.m_Backends, |
| 2469 | isSupported, |
| 2470 | input.GetTensorInfo(), |
| 2471 | outputInfo, |
| 2472 | desc); |
| 2473 | }; |
| 2474 | |
| 2475 | if(IsDynamicTensor(outputInfo)) |
| 2476 | { |
| 2477 | isSupported = AreDynamicTensorsSupported(); |
| 2478 | } |
| 2479 | else |
| 2480 | { |
| 2481 | validateFunc(outputInfo, isSupported); |
| 2482 | } |
| 2483 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2484 | if (!isSupported) |
| 2485 | { |
| 2486 | return false; |
| 2487 | } |
| 2488 | |
| 2489 | IConnectableLayer* layer = data.m_Network->AddSoftmaxLayer(desc); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2490 | if (!layer) |
| 2491 | { |
| 2492 | return Fail("%s: Could not add the SoftmaxLayer", __func__); |
| 2493 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2494 | input.Connect(layer->GetInputSlot(0)); |
| 2495 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2496 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2497 | } |
| 2498 | |
| 2499 | template<typename HalPolicy, |
| 2500 | typename HalOperation = typename HalPolicy::Operation, |
| 2501 | typename HalModel = typename HalPolicy::Model> |
| 2502 | bool ConvertLstm(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2503 | { |
| 2504 | using HalOperand = typename HalPolicy::Operand; |
| 2505 | using HalOperandType = typename HalPolicy::OperandType; |
| 2506 | |
| 2507 | ALOGV("HalPolicy::ConvertLstm()"); |
| 2508 | |
| 2509 | // Inputs: |
| 2510 | // 00: The input: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where |
| 2511 | // “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input. |
| 2512 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2513 | if (!input.IsValid()) |
| 2514 | { |
| 2515 | return Fail("%s: Could not read input 0: input", __func__); |
| 2516 | } |
| 2517 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2518 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 18, model, data); |
| 2519 | if (!outputStateIn.IsValid()) |
| 2520 | { |
| 2521 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 2522 | } |
| 2523 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2524 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 19, model, data); |
| 2525 | if (!cellStateIn.IsValid()) |
| 2526 | { |
| 2527 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 2528 | } |
| 2529 | |
| 2530 | // Get the mandatory input tensors: |
| 2531 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2532 | // [num_units, input_size]. |
| 2533 | const ConstTensorPin inputToForgetWeightsPin = |
| 2534 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 2)); |
| 2535 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2536 | // [num_units, input_size]. |
| 2537 | const ConstTensorPin inputToCellWeightsPin = |
| 2538 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 3)); |
| 2539 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2540 | // [num_units, input_size]. |
| 2541 | const ConstTensorPin inputToOutputWeightsPin = |
| 2542 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 4)); |
| 2543 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2544 | // [num_units, output_size]. |
| 2545 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 2546 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 6)); |
| 2547 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2548 | // [num_units, output_size]. |
| 2549 | const ConstTensorPin recurrentToCellWeightsPin = |
| 2550 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 7)); |
| 2551 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2552 | // [num_units, output_size]. |
| 2553 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 2554 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 8)); |
| 2555 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2556 | const ConstTensorPin forgetGateBiasPin = |
| 2557 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 13, model, data); |
| 2558 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2559 | const ConstTensorPin cellBiasPin = |
| 2560 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 14, model, data); |
| 2561 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2562 | const ConstTensorPin outputGateBiasPin = |
| 2563 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 15, model, data); |
| 2564 | |
| 2565 | if (!inputToForgetWeightsPin.IsValid() || |
| 2566 | !inputToCellWeightsPin.IsValid() || |
| 2567 | !inputToOutputWeightsPin.IsValid() || |
| 2568 | !recurrentToForgetWeightsPin.IsValid() || |
| 2569 | !recurrentToCellWeightsPin.IsValid() || |
| 2570 | !recurrentToOutputWeightsPin.IsValid() || |
| 2571 | !forgetGateBiasPin.IsValid() || |
| 2572 | !cellBiasPin.IsValid() || |
| 2573 | !outputGateBiasPin.IsValid()) |
| 2574 | { |
| 2575 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2576 | } |
| 2577 | |
| 2578 | // Get the optional input tensors: |
| 2579 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2580 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 2581 | const ConstTensorPin inputToInputWeightsPin = |
| 2582 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 1, true)); |
| 2583 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2584 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 2585 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 2586 | const ConstTensorPin recurrentToInputWeightsPin = |
| 2587 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 5, true)); |
| 2588 | // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2589 | const ConstTensorPin cellToInputWeightsPin = |
| 2590 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 9, true)); |
| 2591 | // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2592 | const ConstTensorPin cellToForgetWeightsPin = |
| 2593 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 10, true)); |
| 2594 | // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2595 | const ConstTensorPin cellToOutputWeightsPin = |
| 2596 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 11, true)); |
| 2597 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2598 | const ConstTensorPin inputGateBiasPin = |
| 2599 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2600 | 12, |
| 2601 | model, |
| 2602 | data, |
| 2603 | g_DontPermute, |
| 2604 | nullptr, |
| 2605 | true); |
| 2606 | |
| 2607 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2608 | // [output_size, num_units]. |
| 2609 | const ConstTensorPin projectionWeightsPin = |
| 2610 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 16, true)); |
| 2611 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size]. |
| 2612 | const ConstTensorPin projectionBiasPin = |
| 2613 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2614 | 17, |
| 2615 | model, |
| 2616 | data, |
| 2617 | g_DontPermute, |
| 2618 | nullptr, |
| 2619 | true); |
| 2620 | |
| 2621 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 2622 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 2623 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 2624 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 2625 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 2626 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 2627 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 2628 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 2629 | { |
| 2630 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2631 | } |
| 2632 | |
| 2633 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 2634 | // 20: The activation function: A value indicating the activation function: |
| 2635 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 2636 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 2637 | // If set to 0.0 then clipping is disabled. |
| 2638 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 2639 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 2640 | ActivationFn activation; |
| 2641 | float cellClip; |
| 2642 | float projClip; |
| 2643 | if (!GetInputActivationFunctionFromTensor<HalPolicy>(operation, 20, activation, model, data) || |
| 2644 | !GetInputScalar<HalPolicy>(operation, 21, HalOperandType::FLOAT32, cellClip, model, data) || |
| 2645 | !GetInputScalar<HalPolicy>(operation, 22, HalOperandType::FLOAT32, projClip, model, data)) |
| 2646 | { |
| 2647 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 2648 | } |
| 2649 | |
| 2650 | // Get the normalization tensors |
| 2651 | // 23: The input layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2652 | // Used to rescale normalized inputs to activation at input gate. |
| 2653 | const ConstTensorPin inputLayerNormWeightsPin |
| 2654 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 23, true)); |
| 2655 | |
| 2656 | // 24: The forget layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2657 | // Used to rescale normalized inputs to activation at forget gate. |
| 2658 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 2659 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2660 | 24, |
| 2661 | model, |
| 2662 | data, |
| 2663 | g_DontPermute, |
| 2664 | nullptr, |
| 2665 | true); |
| 2666 | |
| 2667 | // 25: The cell layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2668 | // Used to rescale normalized inputs to activation at cell gate. |
| 2669 | const ConstTensorPin cellLayerNormWeightsPin = |
| 2670 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2671 | 25, |
| 2672 | model, |
| 2673 | data, |
| 2674 | g_DontPermute, |
| 2675 | nullptr, |
| 2676 | true); |
| 2677 | |
| 2678 | // 26: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2679 | // Used to rescale normalized inputs to activation at output gate. |
| 2680 | const ConstTensorPin outputLayerNormWeightsPin = |
| 2681 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2682 | 26, |
| 2683 | model, |
| 2684 | data, |
| 2685 | g_DontPermute, |
| 2686 | nullptr, |
| 2687 | true); |
| 2688 | |
| 2689 | // Outputs: |
| 2690 | // 00: The scratch buffer: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] |
| 2691 | // with CIFG, or [batch_size, num_units * 3] without CIFG. |
| 2692 | const HalOperand* scratchBuffer = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2693 | if (!scratchBuffer) |
| 2694 | { |
| 2695 | return Fail("%s: Could not read output 0: scratchBuffer", __func__); |
| 2696 | } |
| 2697 | // 01: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2698 | const HalOperand* outputStateOut = GetOutputOperand<HalPolicy>(operation, 1, model); |
| 2699 | if (!outputStateOut) |
| 2700 | { |
| 2701 | return Fail("%s: Could not read output 1: outputStateOut", __func__); |
| 2702 | } |
| 2703 | // 02: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2704 | const HalOperand* cellStateOut = GetOutputOperand<HalPolicy>(operation, 2, model); |
| 2705 | if (!cellStateOut) |
| 2706 | { |
| 2707 | return Fail("%s: Could not read output 2: cellStateOut", __func__); |
| 2708 | } |
| 2709 | // 03: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is |
| 2710 | // effectively the same as the current “output state (out)” value. |
| 2711 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 3, model); |
| 2712 | if (!output) |
| 2713 | { |
| 2714 | return Fail("%s: Could not read output 3: output", __func__); |
| 2715 | } |
| 2716 | |
| 2717 | // set the params structure for the AddLstmLayer call |
| 2718 | LstmInputParams params; |
| 2719 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 2720 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 2721 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 2722 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 2723 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 2724 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 2725 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 2726 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 2727 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 2728 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 2729 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 2730 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 2731 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 2732 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 2733 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 2734 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 2735 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 2736 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 2737 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 2738 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 2739 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 2740 | |
| 2741 | // set the layer descriptor |
| 2742 | LstmDescriptor desc; |
| 2743 | desc.m_ActivationFunc = activation; |
| 2744 | desc.m_ClippingThresCell = cellClip; |
| 2745 | desc.m_ClippingThresProj = projClip; |
| 2746 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 2747 | params.m_RecurrentToInputWeights == nullptr || |
| 2748 | params.m_InputGateBias == nullptr); |
| 2749 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 2750 | params.m_CellToOutputWeights != nullptr); |
| 2751 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 2752 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 2753 | params.m_ForgetLayerNormWeights != nullptr || |
| 2754 | params.m_CellLayerNormWeights != nullptr || |
| 2755 | params.m_OutputLayerNormWeights != nullptr); |
| 2756 | |
| 2757 | // validate the optional input groups |
| 2758 | if (desc.m_CifgEnabled && |
| 2759 | (params.m_InputToInputWeights != nullptr || |
| 2760 | params.m_RecurrentToInputWeights != nullptr || |
| 2761 | params.m_InputGateBias != nullptr)) |
| 2762 | { |
| 2763 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 2764 | " and input gate bias must be provided", __func__); |
| 2765 | } |
| 2766 | |
| 2767 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 2768 | { |
| 2769 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 2770 | } |
| 2771 | |
| 2772 | if (desc.m_PeepholeEnabled && |
| 2773 | (params.m_CellToForgetWeights == nullptr || |
| 2774 | params.m_CellToOutputWeights == nullptr || |
| 2775 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 2776 | { |
| 2777 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 2778 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 2779 | } |
| 2780 | |
| 2781 | if (desc.m_LayerNormEnabled && |
| 2782 | (params.m_ForgetLayerNormWeights == nullptr || |
| 2783 | params.m_CellLayerNormWeights == nullptr || |
| 2784 | params.m_OutputLayerNormWeights == nullptr || |
| 2785 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 2786 | { |
| 2787 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 2788 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 2789 | } |
| 2790 | |
| 2791 | // Check if the layer is supported |
| 2792 | // Inputs |
| 2793 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2794 | const TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 2795 | const TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
| 2796 | |
| 2797 | // Outputs |
| 2798 | const TensorInfo& scratchBufferInfo = GetTensorInfoForOperand(*scratchBuffer); |
| 2799 | const TensorInfo& outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); |
| 2800 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 2801 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2802 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2803 | // Basic parameters |
| 2804 | LstmInputParamsInfo paramsInfo; |
| 2805 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 2806 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 2807 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 2808 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 2809 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 2810 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 2811 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 2812 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 2813 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 2814 | |
| 2815 | // Optional parameters |
| 2816 | if (!desc.m_CifgEnabled) |
| 2817 | { |
| 2818 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 2819 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 2820 | if (params.m_CellToInputWeights != nullptr) |
| 2821 | { |
| 2822 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 2823 | } |
| 2824 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 2825 | } |
| 2826 | |
| 2827 | if (desc.m_ProjectionEnabled) |
| 2828 | { |
| 2829 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 2830 | if (params.m_ProjectionBias != nullptr) |
| 2831 | { |
| 2832 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 2833 | } |
| 2834 | } |
| 2835 | |
| 2836 | if (desc.m_PeepholeEnabled) |
| 2837 | { |
| 2838 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 2839 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 2840 | } |
| 2841 | |
| 2842 | if (desc.m_LayerNormEnabled) |
| 2843 | { |
| 2844 | if(!desc.m_CifgEnabled) |
| 2845 | { |
| 2846 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 2847 | } |
| 2848 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 2849 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 2850 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 2851 | } |
| 2852 | |
| 2853 | bool isSupported = false; |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 2854 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2855 | { |
| 2856 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2857 | IsLstmSupported, |
| 2858 | data.m_Backends, |
| 2859 | isSupported, |
| 2860 | inputInfo, |
| 2861 | outputStateInInfo, |
| 2862 | cellStateInInfo, |
| 2863 | scratchBufferInfo, |
| 2864 | outputStateOutInfo, |
| 2865 | cellStateOutInfo, |
| 2866 | outputInfo, |
| 2867 | desc, |
| 2868 | paramsInfo); |
| 2869 | }; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2870 | |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 2871 | bool isDynamic = false; |
| 2872 | if (!IsDynamicTensor(outputStateOutInfo) && |
| 2873 | !IsDynamicTensor(scratchBufferInfo) && |
| 2874 | !IsDynamicTensor(cellStateOutInfo) && |
| 2875 | !IsDynamicTensor(outputInfo)) |
| 2876 | { |
| 2877 | validateFunc(outputInfo, isSupported); |
| 2878 | } |
| 2879 | else |
| 2880 | { |
| 2881 | isDynamic = true; |
| 2882 | isSupported = AreDynamicTensorsSupported(); |
| 2883 | } |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2884 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2885 | if (!isSupported) |
| 2886 | { |
| 2887 | return false; |
| 2888 | } |
| 2889 | |
| 2890 | // Add the layer |
| 2891 | IConnectableLayer* layer = data.m_Network->AddLstmLayer(desc, params, "Lstm"); |
| 2892 | |
| 2893 | input.Connect(layer->GetInputSlot(0)); |
| 2894 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 2895 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 2896 | |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 2897 | if (!isDynamic) |
| 2898 | { |
| 2899 | return ( |
| 2900 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 2901 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data) && |
| 2902 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 2, *layer, 2, model, data) && |
| 2903 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 3, *layer, 3, model, data)); |
| 2904 | } |
| 2905 | else |
| 2906 | { |
| 2907 | return ( |
| 2908 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 2909 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data) && |
| 2910 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 2, *layer, 2, model, data) && |
| 2911 | SetupAndTrackLayerOutputSlot<HalPolicy>( |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 2912 | operation, 3, *layer, 3, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 2913 | } |
| 2914 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2915 | } |
| 2916 | |
| 2917 | template<typename HalPolicy, |
| 2918 | typename HalOperation = typename HalPolicy::Operation, |
| 2919 | typename HalModel = typename HalPolicy::Model> |
| 2920 | bool ConvertTransposeConv2d(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2921 | { |
| 2922 | using HalOperand = typename HalPolicy::Operand; |
| 2923 | using HalOperandType = typename HalPolicy::OperandType; |
| 2924 | |
| 2925 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2926 | |
| 2927 | if (!input.IsValid()) |
| 2928 | { |
| 2929 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2930 | } |
| 2931 | |
| 2932 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2933 | |
| 2934 | if (!output) |
| 2935 | { |
| 2936 | return Fail("%s: Could not read output 0", __func__); |
| 2937 | } |
| 2938 | |
| 2939 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2940 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2941 | |
| 2942 | // ArmNN does not currently support non-fixed weights or bias |
| 2943 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 2944 | const HalOperand* weightsOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
| 2945 | |
| 2946 | if (weightsOperand == nullptr) |
| 2947 | { |
| 2948 | return Fail("%s: Operand is invalid", __func__); |
| 2949 | } |
| 2950 | TransposeConvolution2dDescriptor desc; |
| 2951 | desc.m_DataLayout = DataLayout::NHWC; |
| 2952 | |
| 2953 | // Determine whether padding is implicit or explicit |
| 2954 | bool implicitPadding = operation.inputs.size() == 9; |
| 2955 | |
| 2956 | if (implicitPadding ) |
| 2957 | { |
| 2958 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 8, model, data); |
| 2959 | } |
| 2960 | else |
| 2961 | { |
| 2962 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 10, model, data); |
| 2963 | } |
| 2964 | |
| 2965 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 2966 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 2967 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 2968 | |
| 2969 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
| 2970 | |
| 2971 | // The shape of the weight is [depth_out, filter_height, filter_width, depth_in]. |
| 2972 | // We have to permute it to OIHW if the data layout is NCHW. |
| 2973 | const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ? |
| 2974 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, |
| 2975 | model, data, OHWIToOIHW) : |
| 2976 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 2977 | |
| 2978 | // Bias is a 1D tensor |
| 2979 | const ConstTensorPin biasPin = |
| 2980 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 2981 | |
| 2982 | if (!weightsPin.IsValid()) |
| 2983 | { |
| 2984 | return Fail("%s: Operation has invalid weights", __func__); |
| 2985 | } |
| 2986 | |
| 2987 | if (!biasPin.IsValid()) |
| 2988 | { |
| 2989 | return Fail("%s: Operation has invalid biases", __func__); |
| 2990 | } |
| 2991 | |
| 2992 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 2993 | ConstTensor bias = biasPin.GetConstTensor(); |
| 2994 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 2995 | |
| 2996 | ActivationFn activation; |
| 2997 | |
| 2998 | if (implicitPadding) |
| 2999 | { |
| 3000 | int32_t strideX{0}; |
| 3001 | int32_t strideY{0}; |
| 3002 | int32_t padLeft{0}; |
| 3003 | int32_t padRight{0}; |
| 3004 | int32_t padTop{0}; |
| 3005 | int32_t padBottom{0}; |
| 3006 | |
| 3007 | android::nn::PaddingScheme paddingScheme; |
| 3008 | if (!GetInputPaddingScheme<HalPolicy>(operation, 4, paddingScheme, model, data) || |
| 3009 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, strideX, model, data) || |
| 3010 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, strideY, model, data) || |
| 3011 | !GetInputActivationFunction<HalPolicy>(operation, 7, activation, model, data)) |
| 3012 | { |
| 3013 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 3014 | } |
| 3015 | |
| 3016 | const uint32_t kernelX = weights.GetShape()[widthIndex]; |
| 3017 | const uint32_t kernelY = weights.GetShape()[heightIndex]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3018 | |
Colm Donelan | 8f3d33e | 2020-07-06 15:41:43 +0100 | [diff] [blame] | 3019 | // If output shape has been specified as a parameter then extract it and make it available. |
| 3020 | const HalOperand* outputShapeOperand = GetInputOperand<HalPolicy>(operation, 3, model, false); |
| 3021 | std::vector<int32_t> outputShape; |
| 3022 | if ((outputShapeOperand) && (GetTensorInt32Values<HalPolicy>(*outputShapeOperand, outputShape, model, data))) |
| 3023 | { |
| 3024 | // Change from signed to unsigned int to store in TransposeConvolution2dDescriptor. |
| 3025 | for (int dimension : outputShape) |
| 3026 | { |
| 3027 | desc.m_OutputShape.push_back(static_cast<unsigned int>(dimension)); |
| 3028 | } |
| 3029 | desc.m_OutputShapeEnabled = true; |
| 3030 | } |
| 3031 | |
Finn Williams | 8fe50c6 | 2020-10-09 15:52:57 +0100 | [diff] [blame] | 3032 | uint32_t outputX; |
| 3033 | uint32_t outputY; |
| 3034 | |
| 3035 | if (IsDynamicTensor(outputInfo)) |
| 3036 | { |
| 3037 | if (outputShape.size() == 0) |
| 3038 | { |
| 3039 | return Fail("%s: Padding sizes cannot be inferred", __func__); |
| 3040 | } |
| 3041 | |
| 3042 | outputX = outputShape[widthIndex]; |
| 3043 | outputY = outputShape[heightIndex]; |
| 3044 | } |
| 3045 | else |
| 3046 | { |
| 3047 | outputX = outputInfo.GetShape()[widthIndex]; |
| 3048 | outputY = outputInfo.GetShape()[heightIndex]; |
| 3049 | } |
| 3050 | |
| 3051 | CalcPaddingTransposeConv(outputX, kernelX, strideX, padLeft, padRight, paddingScheme); |
| 3052 | CalcPaddingTransposeConv(outputY, kernelY, strideY, padTop, padBottom, paddingScheme); |
| 3053 | |
| 3054 | // NOTE: The Android NN API allows for negative padding values in TransposeConv2d, |
| 3055 | // but Arm NN only supports values >= 0 |
| 3056 | if (padLeft < 0 || padRight < 0 || padTop < 0 || padBottom < 0) |
| 3057 | { |
| 3058 | return Fail("%s: Negative padding values are not supported", __func__); |
| 3059 | } |
| 3060 | |
Matthew Sloyan | 9b088d9 | 2020-09-14 15:12:55 +0100 | [diff] [blame] | 3061 | desc.m_StrideX = armnn::numeric_cast<uint32_t>(strideX); |
| 3062 | desc.m_StrideY = armnn::numeric_cast<uint32_t>(strideY); |
| 3063 | desc.m_PadLeft = armnn::numeric_cast<uint32_t>(padLeft); |
| 3064 | desc.m_PadRight = armnn::numeric_cast<uint32_t>(padRight); |
| 3065 | desc.m_PadTop = armnn::numeric_cast<uint32_t>(padTop); |
| 3066 | desc.m_PadBottom = armnn::numeric_cast<uint32_t>(padBottom); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3067 | } |
| 3068 | else if (operation.inputs.size() == 11) |
| 3069 | { |
| 3070 | // explicit padding |
| 3071 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 3072 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 3073 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 3074 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 3075 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 3076 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 3077 | !GetInputActivationFunction<HalPolicy>(operation, 9, activation, model, data)) |
| 3078 | { |
| 3079 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 3080 | } |
| 3081 | } |
| 3082 | else |
| 3083 | { |
| 3084 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 3085 | } |
| 3086 | |
| 3087 | desc.m_BiasEnabled = true; |
| 3088 | Optional<TensorInfo> biases(bias.GetInfo()); |
| 3089 | |
| 3090 | bool isSupported = false; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 3091 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3092 | { |
| 3093 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3094 | IsTransposeConvolution2dSupported, |
| 3095 | data.m_Backends, |
| 3096 | isSupported, |
| 3097 | inputInfo, |
| 3098 | outputInfo, |
| 3099 | desc, |
| 3100 | weights.GetInfo(), |
| 3101 | biases); |
| 3102 | }; |
| 3103 | |
| 3104 | if(IsDynamicTensor(outputInfo)) |
| 3105 | { |
| 3106 | isSupported = AreDynamicTensorsSupported(); |
| 3107 | } |
| 3108 | else |
| 3109 | { |
| 3110 | validateFunc(outputInfo, isSupported); |
| 3111 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3112 | if (!isSupported) |
| 3113 | { |
| 3114 | return false; |
| 3115 | } |
| 3116 | |
| 3117 | IConnectableLayer* startLayer = |
| 3118 | data.m_Network->AddTransposeConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
| 3119 | if (!startLayer) |
| 3120 | { |
| 3121 | return Fail("%s: AddTransposeConvolution2dLayer failed", __func__); |
| 3122 | } |
| 3123 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3124 | input.Connect(startLayer->GetInputSlot(0)); |
| 3125 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 3126 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model, |
| 3127 | data, nullptr, validateFunc, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3128 | } |
| 3129 | |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3130 | template<typename HalPolicy, |
| 3131 | typename HalOperation = typename HalPolicy::Operation, |
| 3132 | typename HalModel = typename HalPolicy::Model> |
| 3133 | bool ConvertUnidirectionalSequenceLstm(const HalOperation& operation, |
| 3134 | const HalModel& model, |
| 3135 | ConversionData& data) |
| 3136 | { |
| 3137 | using HalOperand = typename HalPolicy::Operand; |
| 3138 | using HalOperandType = typename HalPolicy::OperandType; |
| 3139 | |
| 3140 | ALOGV("HalPolicy::ConvertUnidirectionalSequenceLstm()"); |
| 3141 | |
| 3142 | // Determine if input OperandType is ANEURALNETWORKS_TENSOR_FLOAT 32 or 16 |
| 3143 | HalOperandType inputType; |
| 3144 | if (!GetOperandType<HalPolicy>(operation, 0, model, inputType)) |
| 3145 | { |
| 3146 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3147 | } |
| 3148 | |
| 3149 | // Inputs: |
| 3150 | // 0: The input: A 3-D tensor of shape: If time-major: [max_time, batch_size, input_size] If batch-major: |
| 3151 | // [batch_size, max_time, input_size] where “max_time” is the number of timesteps (sequence length), “batch_size” |
| 3152 | // corresponds to the batching dimension, and “input_size” is the size of the input. |
| 3153 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 3154 | if (!input.IsValid()) |
| 3155 | { |
| 3156 | return Fail("%s: Could not read input 0: input", __func__); |
| 3157 | } |
| 3158 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [batch_size, output_size]. |
| 3159 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 18, model, data); |
| 3160 | if (!outputStateIn.IsValid()) |
| 3161 | { |
| 3162 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 3163 | } |
| 3164 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [batch_size, num_units]. |
| 3165 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 19, model, data); |
| 3166 | if (!cellStateIn.IsValid()) |
| 3167 | { |
| 3168 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 3169 | } |
| 3170 | |
| 3171 | // Get the mandatory input tensors: |
| 3172 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3173 | // [num_units, input_size]. |
| 3174 | const ConstTensorPin inputToForgetWeightsPin = |
| 3175 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 2)); |
| 3176 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3177 | // [num_units, input_size]. |
| 3178 | const ConstTensorPin inputToCellWeightsPin = |
| 3179 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 3)); |
| 3180 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3181 | // [num_units, input_size]. |
| 3182 | const ConstTensorPin inputToOutputWeightsPin = |
| 3183 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 4)); |
| 3184 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3185 | // [num_units, output_size]. |
| 3186 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 3187 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 6)); |
| 3188 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 3189 | // [num_units, output_size]. |
| 3190 | const ConstTensorPin recurrentToCellWeightsPin = |
| 3191 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 7)); |
| 3192 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3193 | // [num_units, output_size]. |
| 3194 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 3195 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 8)); |
| 3196 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3197 | const ConstTensorPin forgetGateBiasPin = |
| 3198 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 13, model, data); |
| 3199 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3200 | const ConstTensorPin cellBiasPin = |
| 3201 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 14, model, data); |
| 3202 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3203 | const ConstTensorPin outputGateBiasPin = |
| 3204 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 15, model, data); |
| 3205 | |
| 3206 | if (!inputToForgetWeightsPin.IsValid() || |
| 3207 | !inputToCellWeightsPin.IsValid() || |
| 3208 | !inputToOutputWeightsPin.IsValid() || |
| 3209 | !recurrentToForgetWeightsPin.IsValid() || |
| 3210 | !recurrentToCellWeightsPin.IsValid() || |
| 3211 | !recurrentToOutputWeightsPin.IsValid() || |
| 3212 | !forgetGateBiasPin.IsValid() || |
| 3213 | !cellBiasPin.IsValid() || |
| 3214 | !outputGateBiasPin.IsValid()) |
| 3215 | { |
| 3216 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3217 | } |
| 3218 | |
| 3219 | // Get the optional input tensors: |
| 3220 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3221 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 3222 | const ConstTensorPin inputToInputWeightsPin = |
| 3223 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 1, true)); |
| 3224 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3225 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 3226 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 3227 | const ConstTensorPin recurrentToInputWeightsPin = |
| 3228 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 5, true)); |
| 3229 | // 09: The cell-to-input weights: Optional. |
| 3230 | // A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3231 | const ConstTensorPin cellToInputWeightsPin = |
| 3232 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 9, true)); |
| 3233 | // 10: The cell-to-forget weights: Optional. |
| 3234 | // A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3235 | const ConstTensorPin cellToForgetWeightsPin = |
| 3236 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 10, true)); |
| 3237 | // 11: The cell-to-output weights: Optional. |
| 3238 | // A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3239 | const ConstTensorPin cellToOutputWeightsPin = |
| 3240 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 11, true)); |
| 3241 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3242 | const ConstTensorPin inputGateBiasPin = |
| 3243 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3244 | 12, |
| 3245 | model, |
| 3246 | data, |
| 3247 | g_DontPermute, |
| 3248 | nullptr, |
| 3249 | true); |
| 3250 | |
| 3251 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3252 | // [output_size, num_units]. |
| 3253 | const ConstTensorPin projectionWeightsPin = |
| 3254 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 16, true)); |
| 3255 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [output_size]. |
| 3256 | const ConstTensorPin projectionBiasPin = |
| 3257 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3258 | 17, |
| 3259 | model, |
| 3260 | data, |
| 3261 | g_DontPermute, |
| 3262 | nullptr, |
| 3263 | true); |
| 3264 | |
| 3265 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 3266 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 3267 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 3268 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 3269 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 3270 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 3271 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 3272 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 3273 | { |
| 3274 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3275 | } |
| 3276 | |
| 3277 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 3278 | // 20: The activation function: A value indicating the activation function: |
| 3279 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 3280 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 3281 | // If set to 0.0 then clipping is disabled. |
| 3282 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 3283 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 3284 | // Determine data type of input tensor |
| 3285 | ActivationFn activation; |
| 3286 | LstmDescriptor desc; |
| 3287 | |
| 3288 | if (inputType == HalOperandType::TENSOR_FLOAT32) |
| 3289 | { |
| 3290 | float cellClip; |
| 3291 | float projClip; |
| 3292 | |
| 3293 | if (!GetInputActivationFunctionFromTensor<HalPolicy>(operation, 20, activation, model, data) || |
| 3294 | !GetInputScalar<HalPolicy>(operation, 21, HalOperandType::FLOAT32, cellClip, model, data) || |
| 3295 | !GetInputScalar<HalPolicy>(operation, 22, HalOperandType::FLOAT32, projClip, model, data)) |
| 3296 | { |
| 3297 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 3298 | } |
| 3299 | |
| 3300 | desc.m_ClippingThresCell = cellClip; |
| 3301 | desc.m_ClippingThresProj = projClip; |
| 3302 | } |
| 3303 | |
| 3304 | if (inputType == HalOperandType::TENSOR_FLOAT16) |
| 3305 | { |
| 3306 | Half cellClip; |
| 3307 | Half projClip; |
| 3308 | |
| 3309 | if (!GetInputActivationFunctionFromTensor<HalPolicy>(operation, 20, activation, model, data) || |
| 3310 | !GetInputScalar<HalPolicy>(operation, 21, HalOperandType::FLOAT16, cellClip, model, data) || |
| 3311 | !GetInputScalar<HalPolicy>(operation, 22, HalOperandType::FLOAT16, projClip, model, data)) |
| 3312 | { |
| 3313 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 3314 | } |
| 3315 | |
| 3316 | desc.m_ClippingThresCell = cellClip; |
| 3317 | desc.m_ClippingThresProj = projClip; |
| 3318 | } |
| 3319 | |
| 3320 | // Determine if time-major or batch-major. |
| 3321 | // 23: Time-major if true, batch-major if false. |
| 3322 | bool isTimeMajor = GetOptionalBool<HalPolicy>(operation, 23, model, data); |
| 3323 | |
| 3324 | // Get the normalization tensors |
| 3325 | // 24: The input layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3326 | // Used to rescale normalized inputs to activation at input gate. |
| 3327 | const ConstTensorPin inputLayerNormWeightsPin |
| 3328 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 24, true)); |
| 3329 | |
| 3330 | // 25: The forget layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3331 | // Used to rescale normalized inputs to activation at forget gate. |
| 3332 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 3333 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3334 | 25, |
| 3335 | model, |
| 3336 | data, |
| 3337 | g_DontPermute, |
| 3338 | nullptr, |
| 3339 | true); |
| 3340 | |
| 3341 | // 26: The cell layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3342 | // Used to rescale normalized inputs to activation at cell gate. |
| 3343 | const ConstTensorPin cellLayerNormWeightsPin = |
| 3344 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3345 | 26, |
| 3346 | model, |
| 3347 | data, |
| 3348 | g_DontPermute, |
| 3349 | nullptr, |
| 3350 | true); |
| 3351 | |
| 3352 | // 27: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3353 | // Used to rescale normalized inputs to activation at output gate. |
| 3354 | const ConstTensorPin outputLayerNormWeightsPin = |
| 3355 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3356 | 27, |
| 3357 | model, |
| 3358 | data, |
| 3359 | g_DontPermute, |
| 3360 | nullptr, |
| 3361 | true); |
| 3362 | |
| 3363 | // Outputs: |
| 3364 | // 00: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16. Shape: if time-major: |
| 3365 | // [max_time, batch_size, output_size] If batch-major: [batch_size, max_time, output_size] |
| 3366 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 3367 | if (!output) |
| 3368 | { |
| 3369 | return Fail("%s: Could not read output: ", __func__); |
| 3370 | } |
| 3371 | |
| 3372 | // |
| 3373 | // 01 & 02: |
| 3374 | // hiddenStateOut and cellStateOut are not currently supported by our android versioning. |
| 3375 | // |
| 3376 | |
| 3377 | // set the params structure for the AddLstmLayer call |
| 3378 | LstmInputParams params; |
| 3379 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 3380 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 3381 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 3382 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 3383 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 3384 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 3385 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 3386 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 3387 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 3388 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 3389 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 3390 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 3391 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 3392 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 3393 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 3394 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 3395 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 3396 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 3397 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 3398 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 3399 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 3400 | |
| 3401 | // set the layer descriptor |
| 3402 | desc.m_ActivationFunc = activation; |
| 3403 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 3404 | params.m_RecurrentToInputWeights == nullptr || |
| 3405 | params.m_InputGateBias == nullptr); |
| 3406 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 3407 | params.m_CellToOutputWeights != nullptr); |
| 3408 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 3409 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 3410 | params.m_ForgetLayerNormWeights != nullptr || |
| 3411 | params.m_CellLayerNormWeights != nullptr || |
| 3412 | params.m_OutputLayerNormWeights != nullptr); |
| 3413 | desc.m_TimeMajor = isTimeMajor; |
| 3414 | |
| 3415 | // validate the optional input groups |
| 3416 | if (desc.m_CifgEnabled && |
| 3417 | (params.m_InputToInputWeights != nullptr || |
| 3418 | params.m_RecurrentToInputWeights != nullptr || |
| 3419 | params.m_InputGateBias != nullptr)) |
| 3420 | { |
| 3421 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 3422 | " and input gate bias must be provided", __func__); |
| 3423 | } |
| 3424 | |
| 3425 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 3426 | { |
| 3427 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 3428 | } |
| 3429 | |
| 3430 | if (desc.m_PeepholeEnabled && |
| 3431 | (params.m_CellToForgetWeights == nullptr || |
| 3432 | params.m_CellToOutputWeights == nullptr || |
| 3433 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 3434 | { |
| 3435 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 3436 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 3437 | } |
| 3438 | |
| 3439 | if (desc.m_LayerNormEnabled && |
| 3440 | (params.m_ForgetLayerNormWeights == nullptr || |
| 3441 | params.m_CellLayerNormWeights == nullptr || |
| 3442 | params.m_OutputLayerNormWeights == nullptr || |
| 3443 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 3444 | { |
| 3445 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 3446 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 3447 | } |
| 3448 | |
| 3449 | // Check if the layer is supported |
| 3450 | // Inputs |
| 3451 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3452 | const TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 3453 | const TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
| 3454 | |
| 3455 | // Outputs |
| 3456 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3457 | |
| 3458 | // Basic parameters |
| 3459 | LstmInputParamsInfo paramsInfo; |
| 3460 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 3461 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 3462 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 3463 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 3464 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 3465 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 3466 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 3467 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 3468 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 3469 | |
| 3470 | // Optional parameters |
| 3471 | if (!desc.m_CifgEnabled) |
| 3472 | { |
| 3473 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 3474 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 3475 | if (params.m_CellToInputWeights != nullptr) |
| 3476 | { |
| 3477 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 3478 | } |
| 3479 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 3480 | } |
| 3481 | |
| 3482 | if (desc.m_ProjectionEnabled) |
| 3483 | { |
| 3484 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 3485 | if (params.m_ProjectionBias != nullptr) |
| 3486 | { |
| 3487 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 3488 | } |
| 3489 | } |
| 3490 | |
| 3491 | if (desc.m_PeepholeEnabled) |
| 3492 | { |
| 3493 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 3494 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 3495 | } |
| 3496 | |
| 3497 | if (desc.m_LayerNormEnabled) |
| 3498 | { |
| 3499 | if(!desc.m_CifgEnabled) |
| 3500 | { |
| 3501 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 3502 | } |
| 3503 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 3504 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 3505 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 3506 | } |
| 3507 | |
| 3508 | auto hiddenStateOutInfo = EmptyOptional(); |
| 3509 | auto cellStateOutInfo = EmptyOptional(); |
| 3510 | |
| 3511 | bool isSupported = false; |
| 3512 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3513 | { |
| 3514 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3515 | IsUnidirectionalSequenceLstmSupported, |
| 3516 | data.m_Backends, |
| 3517 | isSupported, |
| 3518 | inputInfo, |
| 3519 | outputStateInInfo, |
| 3520 | cellStateInInfo, |
| 3521 | outputInfo, |
| 3522 | hiddenStateOutInfo, |
| 3523 | cellStateOutInfo, |
| 3524 | desc, |
| 3525 | paramsInfo); |
| 3526 | }; |
| 3527 | |
| 3528 | bool isDynamic = false; |
| 3529 | if (!IsDynamicTensor(outputInfo)) |
| 3530 | { |
| 3531 | validateFunc(outputInfo, isSupported); |
| 3532 | } |
| 3533 | else |
| 3534 | { |
| 3535 | isDynamic = true; |
| 3536 | isSupported = AreDynamicTensorsSupported(); |
| 3537 | } |
| 3538 | |
| 3539 | if (!isSupported) |
| 3540 | { |
| 3541 | return false; |
| 3542 | } |
| 3543 | |
| 3544 | // Add the layer |
| 3545 | IConnectableLayer* layer = data.m_Network->AddUnidirectionalSequenceLstmLayer(desc, |
| 3546 | params, |
| 3547 | "UnidirectionalSequenceLstm"); |
| 3548 | |
| 3549 | input.Connect(layer->GetInputSlot(0)); |
| 3550 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 3551 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 3552 | |
| 3553 | if (!isDynamic) |
| 3554 | { |
| 3555 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data)); |
| 3556 | } |
| 3557 | else |
| 3558 | { |
| 3559 | return (SetupAndTrackLayerOutputSlot<HalPolicy>( |
| 3560 | operation, 0, *layer, 0, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
| 3561 | } |
| 3562 | } |
| 3563 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3564 | } // armnn_driver namespace |