Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1 | // |
Teresa Charlin | ee5872d | 2021-12-03 16:07:42 +0000 | [diff] [blame] | 2 | // Copyright © 2020-2023 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, |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 26 | typename HalOperation = typename HalPolicy::Operation, |
| 27 | typename HalModel = typename HalPolicy::Model> |
| 28 | bool IsWeightsValid(const HalOperation& operation, |
| 29 | uint32_t inputIndex, |
Kevin May | f762b83 | 2023-08-18 14:01:51 +0100 | [diff] [blame] | 30 | const HalModel& model, |
| 31 | const bool isOptional = true) |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 32 | { |
| 33 | using HalOperand = typename HalPolicy::Operand; |
| 34 | using HalOperandLifeTime = typename HalPolicy::OperandLifeTime; |
| 35 | const HalOperand* operand = GetInputOperand<HalPolicy>(operation, inputIndex, model); |
| 36 | if (!operand) |
| 37 | { |
| 38 | Fail("%s: failed to get input operand %i", __func__, inputIndex); |
| 39 | return false; |
| 40 | } |
| 41 | |
Kevin May | f762b83 | 2023-08-18 14:01:51 +0100 | [diff] [blame] | 42 | // If the operand is not an optional operand it cannot have a NO_VALUE lifetime |
| 43 | if (!isOptional && operand->lifetime == HalOperandLifeTime::NO_VALUE) |
| 44 | { |
| 45 | return false; |
| 46 | } |
| 47 | |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 48 | if (operand->lifetime != HalOperandLifeTime::CONSTANT_COPY |
| 49 | && operand->lifetime != HalOperandLifeTime::CONSTANT_REFERENCE |
| 50 | && operand->lifetime != HalOperandLifeTime::NO_VALUE) |
| 51 | { |
| 52 | return false; |
| 53 | } |
Kevin May | f762b83 | 2023-08-18 14:01:51 +0100 | [diff] [blame] | 54 | |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 55 | return true; |
| 56 | } |
| 57 | |
| 58 | template<typename HalPolicy, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 59 | typename HalOperation = typename HalPolicy::Operation, |
| 60 | typename HalModel = typename HalPolicy::Model> |
| 61 | bool IsQSymmDequantizeForWeights(const HalOperation& operation, const HalModel& model) |
| 62 | { |
| 63 | using HalOperand = typename HalPolicy::Operand; |
| 64 | using HalOperationType = typename HalPolicy::OperationType; |
| 65 | |
| 66 | const HalOperand* operand = GetInputOperand<HalPolicy>(operation, 0, model); |
| 67 | if (!operand) |
| 68 | { |
| 69 | return false; |
| 70 | } |
| 71 | |
| 72 | if(!IsQSymm8(*operand)) |
| 73 | { |
| 74 | // Only QSymm8 weights are dequantized on the fly by the driver |
| 75 | return false; |
| 76 | } |
| 77 | |
| 78 | if (!IsOperandConstant<HalPolicy>(*operand)) |
| 79 | { |
| 80 | // Non-const input is not accepted for weights |
| 81 | return false; |
| 82 | } |
| 83 | |
| 84 | // Iterate through all the operations and find the operation feeding from the Dequantize output |
| 85 | const size_t outputIndex = operation.outputs[0]; |
| 86 | for (uint32_t operationIdx = 0; operationIdx < getMainModel(model).operations.size(); ++operationIdx) |
| 87 | { |
| 88 | const auto& operationIt = getMainModel(model).operations[operationIdx]; |
| 89 | switch (operationIt.type) |
| 90 | { |
| 91 | case HalOperationType::FULLY_CONNECTED: |
| 92 | if (outputIndex == operationIt.inputs[1]) // Weights are bound to slot 1 |
| 93 | { |
| 94 | // If the output is going into the FC weights return true |
| 95 | return true; |
| 96 | } |
| 97 | break; |
| 98 | case HalOperationType::LSTM: |
| 99 | for (size_t k = 0; k < operationIt.inputs.size(); ++k) |
| 100 | { |
| 101 | if (outputIndex == operationIt.inputs[k]) |
| 102 | { |
| 103 | // If the output is going into the LSTM weights return true |
| 104 | return true; |
| 105 | } |
| 106 | } |
| 107 | break; |
| 108 | default: |
| 109 | break; |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | return false; |
| 114 | } |
| 115 | |
| 116 | template<typename HalPolicy, |
| 117 | typename HalOperation = typename HalPolicy::Operation, |
| 118 | typename HalModel = typename HalPolicy::Model> |
| 119 | bool SetupAndTrackLayerOutputSlotAndOverrideTensorInfo(const HalOperation& operation, |
| 120 | uint32_t operationOutputIndex, |
| 121 | armnn::IConnectableLayer& layer, |
| 122 | uint32_t layerOutputIndex, |
| 123 | const HalModel& model, |
| 124 | ConversionData& data, |
| 125 | const armnn::TensorInfo tensor_info) |
| 126 | { |
| 127 | using HalOperand = typename HalPolicy::Operand; |
| 128 | |
| 129 | const HalOperand* outputOperand = GetOutputOperand<HalPolicy>(operation, operationOutputIndex, model); |
| 130 | if ((outputOperand == nullptr) || (operationOutputIndex >= layer.GetNumOutputSlots())) |
| 131 | { |
| 132 | return false; |
| 133 | } |
| 134 | |
| 135 | armnn::IOutputSlot& outputSlot = layer.GetOutputSlot(layerOutputIndex); |
| 136 | |
| 137 | const uint32_t operandIndex = operation.outputs[operationOutputIndex]; |
| 138 | data.m_OutputSlotForOperand[operandIndex] = &outputSlot; |
| 139 | |
| 140 | outputSlot.SetTensorInfo(tensor_info); |
| 141 | |
| 142 | return true; |
| 143 | } |
| 144 | |
| 145 | template<typename HalPolicy, |
Sadik Armagan | 92b5fd1 | 2021-04-26 09:52:06 +0100 | [diff] [blame] | 146 | typename HalOperation = typename HalPolicy::Operation, |
| 147 | typename HalModel = typename HalPolicy::Model> |
| 148 | bool ConvertCast(const HalOperation& operation, |
| 149 | const HalModel& model, |
| 150 | ConversionData& data) |
| 151 | { |
| 152 | using HalOperand = typename HalPolicy::Operand; |
| 153 | |
| 154 | ALOGV("HalPolicy::ConvertCast()"); |
| 155 | |
| 156 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 157 | |
| 158 | if (!input.IsValid()) |
| 159 | { |
| 160 | return Fail("%s: Operation has invalid inputs", __func__); |
| 161 | } |
| 162 | |
| 163 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 164 | if (!output) |
| 165 | { |
| 166 | return Fail("%s: Could not read output 0", __func__); |
| 167 | } |
| 168 | |
| 169 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 170 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 171 | |
| 172 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 173 | armnn::BackendId setBackend; |
Sadik Armagan | 92b5fd1 | 2021-04-26 09:52:06 +0100 | [diff] [blame] | 174 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 175 | { |
| 176 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 177 | IsCastSupported, |
| 178 | data.m_Backends, |
| 179 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 180 | setBackend, |
Sadik Armagan | 92b5fd1 | 2021-04-26 09:52:06 +0100 | [diff] [blame] | 181 | inputInfo, |
| 182 | outputInfo); |
| 183 | }; |
| 184 | |
| 185 | if(!IsDynamicTensor(outputInfo)) |
| 186 | { |
| 187 | validateFunc(outputInfo, isSupported); |
| 188 | } |
| 189 | else |
| 190 | { |
| 191 | isSupported = AreDynamicTensorsSupported(); |
| 192 | } |
| 193 | |
| 194 | if (!isSupported) |
| 195 | { |
| 196 | return false; |
| 197 | } |
| 198 | |
| 199 | IConnectableLayer* layer = data.m_Network->AddCastLayer(); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 200 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 201 | if (!layer) |
| 202 | { |
| 203 | return Fail("%s: Could not add the CastLayer", __func__); |
| 204 | } |
Sadik Armagan | 92b5fd1 | 2021-04-26 09:52:06 +0100 | [diff] [blame] | 205 | input.Connect(layer->GetInputSlot(0)); |
| 206 | |
| 207 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 208 | } |
| 209 | |
| 210 | template<typename HalPolicy, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 211 | typename HalOperation = typename HalPolicy::Operation, |
| 212 | typename HalModel = typename HalPolicy::Model> |
Teresa Charlin | 7f5b51e | 2021-09-01 14:19:38 +0100 | [diff] [blame] | 213 | bool ConvertChannelShuffle(const HalOperation& operation, |
| 214 | const HalModel& model, |
| 215 | ConversionData& data) |
| 216 | { |
| 217 | using HalOperand = typename HalPolicy::Operand; |
| 218 | using HalOperandType = typename HalPolicy::OperandType; |
| 219 | |
| 220 | ALOGV("HalPolicy::ConvertChannelShuffle()"); |
| 221 | |
| 222 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 223 | if (!input.IsValid()) |
| 224 | { |
| 225 | return Fail("%s: Operation has invalid inputs", __func__); |
| 226 | } |
| 227 | auto inputDimensions = static_cast<int32_t>(input.GetTensorInfo().GetNumDimensions()); |
| 228 | |
| 229 | ChannelShuffleDescriptor descriptor; |
| 230 | |
| 231 | int32_t groups; |
| 232 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, groups, model, data)) |
| 233 | { |
| 234 | return Fail("%s: Operation has invalid or unsupported number of groups operand", __func__); |
| 235 | } |
| 236 | descriptor.m_NumGroups = static_cast<uint32_t>(groups); |
| 237 | |
| 238 | int32_t axis; |
| 239 | if (!GetInputScalar<HalPolicy>(operation, 2, HalOperandType::INT32, axis, model, data)) |
| 240 | { |
| 241 | return Fail("%s: Operation has invalid or unsupported dimension channel shuffle operand", __func__); |
| 242 | } |
| 243 | if (((axis < -inputDimensions) && (axis < 0)) || ((axis >= inputDimensions) && (axis > 0))) |
| 244 | { |
| 245 | return Fail("%s: Operation has invalid dimension: %d. It is out of bounds [-%d, %d))", __func__, axis, |
| 246 | inputDimensions, inputDimensions); |
| 247 | } |
| 248 | int positiveAxis = (axis < 0) ? inputDimensions + axis : axis; |
| 249 | descriptor.m_Axis = static_cast<uint32_t>(positiveAxis); |
| 250 | |
| 251 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 252 | if (!output) |
| 253 | { |
| 254 | return Fail("%s: Could not read output 0", __func__); |
| 255 | } |
| 256 | |
| 257 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 258 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 259 | |
| 260 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 261 | armnn::BackendId setBackend; |
Teresa Charlin | 7f5b51e | 2021-09-01 14:19:38 +0100 | [diff] [blame] | 262 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 263 | { |
| 264 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 265 | IsChannelShuffleSupported, |
| 266 | data.m_Backends, |
| 267 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 268 | setBackend, |
Teresa Charlin | 7f5b51e | 2021-09-01 14:19:38 +0100 | [diff] [blame] | 269 | inputInfo, |
| 270 | outputInfo, |
| 271 | descriptor); |
| 272 | }; |
| 273 | |
| 274 | if(!IsDynamicTensor(outputInfo)) |
| 275 | { |
| 276 | validateFunc(outputInfo, isSupported); |
| 277 | } |
| 278 | else |
| 279 | { |
| 280 | isSupported = AreDynamicTensorsSupported(); |
| 281 | } |
| 282 | |
| 283 | if (!isSupported) |
| 284 | { |
| 285 | return false; |
| 286 | } |
| 287 | |
| 288 | IConnectableLayer* layer = data.m_Network->AddChannelShuffleLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 289 | layer->SetBackendId(setBackend); |
Teresa Charlin | 7f5b51e | 2021-09-01 14:19:38 +0100 | [diff] [blame] | 290 | assert(layer != nullptr); |
| 291 | input.Connect(layer->GetInputSlot(0)); |
| 292 | |
| 293 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 294 | } |
| 295 | |
| 296 | template<typename HalPolicy, |
| 297 | typename HalOperation = typename HalPolicy::Operation, |
| 298 | typename HalModel = typename HalPolicy::Model> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 299 | bool ConvertComparison_1_2(const HalOperation& operation, |
| 300 | const HalModel& model, |
| 301 | ConversionData& data, |
| 302 | ComparisonOperation comparisonOperation) |
| 303 | { |
| 304 | using HalOperand = typename HalPolicy::Operand; |
| 305 | |
| 306 | ALOGV("HalPolicy::ConvertComparison()"); |
| 307 | ALOGV("comparisonOperation = %s", GetComparisonOperationAsCString(comparisonOperation)); |
| 308 | |
| 309 | LayerInputHandle input0 = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 310 | LayerInputHandle input1 = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 311 | |
| 312 | if (!(input0.IsValid() && input1.IsValid())) |
| 313 | { |
| 314 | return Fail("%s: Operation has invalid inputs", __func__); |
| 315 | } |
| 316 | |
| 317 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 318 | if (!output) |
| 319 | { |
| 320 | return Fail("%s: Could not read output 0", __func__); |
| 321 | } |
| 322 | |
| 323 | const TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 324 | const TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 325 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 326 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 327 | ComparisonDescriptor descriptor(comparisonOperation); |
| 328 | |
| 329 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 330 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 331 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 332 | { |
| 333 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 334 | IsComparisonSupported, |
| 335 | data.m_Backends, |
| 336 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 337 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 338 | inputInfo0, |
| 339 | inputInfo1, |
| 340 | outputInfo, |
| 341 | descriptor); |
| 342 | |
| 343 | }; |
| 344 | |
| 345 | if(!IsDynamicTensor(outputInfo)) |
| 346 | { |
| 347 | validateFunc(outputInfo, isSupported); |
| 348 | } |
| 349 | else |
| 350 | { |
| 351 | isSupported = AreDynamicTensorsSupported(); |
| 352 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 353 | |
| 354 | if (!isSupported) |
| 355 | { |
| 356 | return false; |
| 357 | } |
| 358 | |
| 359 | IConnectableLayer* layer = data.m_Network->AddComparisonLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 360 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 361 | if (!layer) |
| 362 | { |
| 363 | return Fail("%s: Could not add the ComparisonLayer", __func__); |
| 364 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 365 | |
| 366 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 367 | if (!isReshapeSupported) |
| 368 | { |
| 369 | return false; |
| 370 | } |
| 371 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 372 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 373 | } |
| 374 | |
| 375 | template<typename HalPolicy, |
| 376 | typename HalOperation = typename HalPolicy::Operation, |
| 377 | typename HalModel = typename HalPolicy::Model> |
| 378 | bool ConvertConv2d_1_2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 379 | { |
| 380 | |
| 381 | using HalOperand = typename HalPolicy::Operand; |
| 382 | using HalOperandType = typename HalPolicy::OperandType; |
| 383 | |
| 384 | ALOGV("HalPolicy::ConvertConv2d_1_2()"); |
| 385 | |
| 386 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 387 | if (!input.IsValid()) |
| 388 | { |
| 389 | return Fail("%s: Operation has invalid inputs", __func__); |
| 390 | } |
| 391 | |
| 392 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 393 | if (!output) |
| 394 | { |
| 395 | return Fail("%s: Could not read output 0", __func__); |
| 396 | } |
| 397 | |
| 398 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 399 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 400 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 401 | Convolution2dDescriptor desc; |
| 402 | desc.m_DataLayout = DataLayout::NHWC; |
| 403 | |
| 404 | // Determine whether padding is implicit or explicit |
| 405 | bool implicitPadding = operation.inputs.size() == 7 || |
| 406 | (operation.inputs.size() >= 8 && |
| 407 | GetInputOperand<HalPolicy>(operation, 7, model)->type == HalOperandType::BOOL); |
| 408 | |
| 409 | if (implicitPadding) |
| 410 | { |
| 411 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 7, model, data); |
| 412 | } |
| 413 | else if (operation.inputs.size() >= 10) |
| 414 | { |
| 415 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 10, model, data); |
| 416 | } |
| 417 | |
| 418 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
| 419 | |
| 420 | // ArmNN does not currently support non-fixed weights or bias |
| 421 | // The NNAPI filter is always OHWI [depth_out, filter_height, filter_width, depth_in] but ArmNN expects the |
| 422 | // filter's height and width indices to match the input's height and width indices so we permute it to OIHW if |
| 423 | // the DataLayout is NCHW |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 424 | |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 425 | |
Kevin May | f762b83 | 2023-08-18 14:01:51 +0100 | [diff] [blame] | 426 | if (!IsWeightsValid<HalPolicy>(operation, 1, model, false) && desc.m_DataLayout == DataLayout::NCHW) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 427 | { |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 428 | return Fail("%s: Operation has unsupported weights HalOperandLifeTime", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 429 | } |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 430 | LayerInputHandle weightsInput = (desc.m_DataLayout == DataLayout::NCHW) ? |
| 431 | ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data, OHWIToOIHW) : |
| 432 | ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 433 | |
| 434 | if (!weightsInput.IsValid()) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 435 | { |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 436 | return Fail("%s: Operation has invalid inputs", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 437 | } |
| 438 | |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 439 | LayerInputHandle biasInput = ConvertToLayerInputHandle<HalPolicy>(operation, 2, model, data); // 1D |
| 440 | if (!biasInput.IsValid()) |
| 441 | { |
| 442 | return Fail("%s: Operation has invalid inputs", __func__); |
| 443 | } |
| 444 | |
| 445 | biasInput.SanitizeQuantizationScale(weightsInput, input); |
| 446 | armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo(); |
| 447 | armnn::TensorInfo biasInfo = biasInput.GetTensorInfo(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 448 | |
| 449 | ActivationFn activation; |
| 450 | |
| 451 | if (implicitPadding) |
| 452 | { |
| 453 | android::nn::PaddingScheme paddingScheme; |
| 454 | if (!GetInputPaddingScheme<HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 455 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 456 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 457 | !GetInputActivationFunction<HalPolicy>(operation, 6, activation, model, data) || |
| 458 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 8, desc, model, data)) |
| 459 | { |
| 460 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 461 | } |
| 462 | |
| 463 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 464 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 465 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 466 | const uint32_t kernelX = weightsInfo.GetShape()[widthIndex]; |
| 467 | const uint32_t kernelY = weightsInfo.GetShape()[heightIndex]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 468 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 469 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 470 | |
| 471 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 472 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 473 | |
| 474 | } |
| 475 | else if (operation.inputs.size() >= 10) |
| 476 | { |
| 477 | // explicit padding |
| 478 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 479 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 480 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 481 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 482 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 483 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 484 | !GetInputActivationFunction<HalPolicy>(operation, 9, activation, model, data) || |
| 485 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 11, desc, model, data)) |
| 486 | { |
| 487 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 488 | } |
| 489 | } |
| 490 | else |
| 491 | { |
| 492 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 493 | } |
| 494 | |
| 495 | desc.m_BiasEnabled = true; |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 496 | Optional<TensorInfo> biases(biasInfo); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 497 | |
| 498 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 499 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 500 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 501 | { |
| 502 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 503 | IsConvolution2dSupported, |
| 504 | data.m_Backends, |
| 505 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 506 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 507 | inputInfo, |
| 508 | outputInfo, |
| 509 | desc, |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 510 | weightsInfo, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 511 | biases); |
| 512 | }; |
| 513 | |
| 514 | if(!IsDynamicTensor(outputInfo)) |
| 515 | { |
| 516 | validateFunc(outputInfo, isSupported); |
| 517 | } |
| 518 | else |
| 519 | { |
| 520 | isSupported = AreDynamicTensorsSupported(); |
| 521 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 522 | |
| 523 | if (!isSupported) |
| 524 | { |
| 525 | return false; |
| 526 | } |
| 527 | |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 528 | armnn::IConnectableLayer* startLayer = data.m_Network->AddConvolution2dLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 529 | startLayer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 530 | |
| 531 | if (!startLayer) |
| 532 | { |
| 533 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 534 | } |
| 535 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 536 | input.Connect(startLayer->GetInputSlot(0)); |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 537 | weightsInput.Connect(startLayer->GetInputSlot(1)); |
| 538 | biasInput.Connect(startLayer->GetInputSlot(2)); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 539 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 540 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model, |
| 541 | data, nullptr, validateFunc, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 542 | } |
| 543 | |
| 544 | template<typename HalPolicy, |
| 545 | typename HalOperation = typename HalPolicy::Operation, |
| 546 | typename HalModel = typename HalPolicy::Model> |
| 547 | bool ConvertDepthwiseConv2d_1_2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 548 | { |
| 549 | using HalOperand = typename HalPolicy::Operand; |
| 550 | using HalOperandType = typename HalPolicy::OperandType; |
| 551 | |
| 552 | ALOGV("HalPolicy::ConvertDepthwiseConv2d_1_2()"); |
| 553 | |
| 554 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 555 | |
| 556 | if (!input.IsValid()) |
| 557 | { |
| 558 | return Fail("%s: Operation has invalid inputs", __func__); |
| 559 | } |
| 560 | |
| 561 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 562 | |
| 563 | if (!output) |
| 564 | { |
| 565 | return Fail("%s: Could not read output 0", __func__); |
| 566 | } |
| 567 | |
| 568 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 569 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 570 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 571 | // ArmNN does not currently support non-fixed weights or bias |
Kevin May | f762b83 | 2023-08-18 14:01:51 +0100 | [diff] [blame] | 572 | if (!IsWeightsValid<HalPolicy>(operation, 1, model, false)) |
| 573 | { |
| 574 | return Fail("%s: This Operation has unsupported weights HalOperandLifeTime", __func__); |
| 575 | } |
| 576 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 577 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 578 | const HalOperand* weightsOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
Kevin May | f762b83 | 2023-08-18 14:01:51 +0100 | [diff] [blame] | 579 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 580 | if (weightsOperand->dimensions[0] != 1) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 581 | { |
| 582 | return Fail("%s: Invalid weights; for depthwise convolution, dimension 0 must be 1 but it is %i", |
| 583 | __func__, weightsOperand->dimensions[0] ); |
| 584 | } |
| 585 | |
| 586 | DepthwiseConvolution2dDescriptor desc; |
| 587 | desc.m_DataLayout = DataLayout::NHWC; |
| 588 | |
| 589 | // Determine whether padding is implicit or explicit |
| 590 | bool implicitPadding = operation.inputs.size() == 8 || |
| 591 | (operation.inputs.size() >= 9 && |
| 592 | GetInputOperand<HalPolicy>(operation, 8, model)->type == HalOperandType::BOOL); |
| 593 | |
| 594 | // Look ahead to find the optional DataLayout, if present |
| 595 | const uint32_t dataLayoutFlagIndex = implicitPadding ? 8 : 11; |
| 596 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, dataLayoutFlagIndex, model, data); |
| 597 | |
| 598 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 599 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 600 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 601 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 602 | LayerInputHandle weightsInput = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 603 | if (!weightsInput.IsValid()) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 604 | { |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 605 | return Fail("%s: Operation has invalid inputs", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 606 | } |
| 607 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 608 | const HalOperand* biasOperand = GetInputOperand<HalPolicy>(operation, 2, model); |
| 609 | if (!biasOperand) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 610 | { |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 611 | return Fail("%s: Could not read bias", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 612 | } |
| 613 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 614 | LayerInputHandle biasInput = ConvertToLayerInputHandle<HalPolicy>(operation, 2, model, data); // 1D |
| 615 | if (!biasInput.IsValid()) |
| 616 | { |
| 617 | return Fail("%s: Operation has invalid inputs", __func__); |
| 618 | } |
| 619 | |
| 620 | biasInput.SanitizeQuantizationScale(weightsInput, input); |
| 621 | armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo(); |
| 622 | armnn::TensorInfo biasInfo = biasInput.GetTensorInfo(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 623 | |
| 624 | ActivationFn activation; |
| 625 | |
| 626 | if (implicitPadding) |
| 627 | { |
| 628 | android::nn::PaddingScheme paddingScheme; |
| 629 | if (!GetInputPaddingScheme<HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 630 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 631 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 632 | !GetInputActivationFunction<HalPolicy>(operation, 7, activation, model, data) || |
| 633 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 9, desc, model, data)) |
| 634 | { |
| 635 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 636 | } |
| 637 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 638 | const uint32_t kernelX = weightsInfo.GetShape()[2]; |
| 639 | const uint32_t kernelY = weightsInfo.GetShape()[1]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 640 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 641 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 642 | |
| 643 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 644 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 645 | } |
| 646 | else if (operation.inputs.size() >= 11) |
| 647 | { |
| 648 | // explicit padding |
| 649 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 650 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 651 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 652 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 653 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 654 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 655 | !GetInputActivationFunction<HalPolicy>(operation, 10, activation, model, data) || |
| 656 | !GetOptionalConvolutionDilationParams<HalPolicy>(operation, 12, desc, model, data)) |
| 657 | { |
| 658 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 659 | } |
| 660 | } |
| 661 | else |
| 662 | { |
| 663 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 664 | } |
| 665 | |
| 666 | desc.m_BiasEnabled = true; |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 667 | Optional<TensorInfo> biases(biasInfo); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 668 | |
| 669 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 670 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 671 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 672 | { |
| 673 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 674 | IsDepthwiseConvolutionSupported, |
| 675 | data.m_Backends, |
| 676 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 677 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 678 | inputInfo, |
| 679 | outputInfo, |
| 680 | desc, |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 681 | weightsInfo, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 682 | biases); |
| 683 | }; |
| 684 | |
| 685 | if(!IsDynamicTensor(outputInfo)) |
| 686 | { |
| 687 | validateFunc(outputInfo, isSupported); |
| 688 | } |
| 689 | else |
| 690 | { |
| 691 | isSupported = AreDynamicTensorsSupported(); |
| 692 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 693 | |
| 694 | if (!isSupported) |
| 695 | { |
| 696 | return false; |
| 697 | } |
| 698 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 699 | armnn::IConnectableLayer* startLayer = data.m_Network->AddDepthwiseConvolution2dLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 700 | startLayer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 701 | |
| 702 | if (!startLayer) |
| 703 | { |
| 704 | return Fail("%s: AddDepthwiseConvolution2dLayer failed", __func__); |
| 705 | } |
| 706 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 707 | input.Connect(startLayer->GetInputSlot(0)); |
| 708 | |
Cathal Corbett | 915f2a7 | 2022-04-15 14:12:08 +0100 | [diff] [blame] | 709 | // Connect weights and bias inputs |
| 710 | weightsInput.Connect(startLayer->GetInputSlot(1)); |
| 711 | biasInput.Connect(startLayer->GetInputSlot(2)); |
| 712 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 713 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model, |
| 714 | data, nullptr, validateFunc, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 715 | } |
| 716 | |
| 717 | template<typename HalPolicy, |
| 718 | typename HalOperation = typename HalPolicy::Operation, |
| 719 | typename HalModel = typename HalPolicy::Model> |
| 720 | bool ConvertDequantize_1_2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 721 | { |
| 722 | ALOGV("HalPolicy::ConvertDequantize()"); |
| 723 | |
| 724 | if (IsQSymmDequantizeForWeights<HalPolicy>(operation, model)) |
| 725 | { |
| 726 | // NOTE: QSymm8 weights are dequantized internally by the driver, |
| 727 | // therefore this type of Dequantize is implicitly supported |
| 728 | return true; |
| 729 | } |
| 730 | |
| 731 | return ::ConvertDequantize<HalPolicy>(operation, model, data); |
| 732 | } |
| 733 | |
| 734 | template<typename HalPolicy, |
| 735 | typename HalOperation = typename HalPolicy::Operation, |
| 736 | typename HalModel = typename HalPolicy::Model> |
| 737 | bool ConvertElementwiseUnary(const HalOperation& operation, |
| 738 | const HalModel& model, |
| 739 | ConversionData& data, |
| 740 | UnaryOperation unaryOperation) |
| 741 | { |
| 742 | using HalOperand = typename HalPolicy::Operand; |
| 743 | |
| 744 | ALOGV("HalPolicy::ConvertElementwiseUnary()"); |
| 745 | ALOGV("unaryOperation = %s", GetUnaryOperationAsCString(unaryOperation)); |
| 746 | |
| 747 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 748 | |
| 749 | if (!input.IsValid()) |
| 750 | { |
| 751 | return Fail("%s: Operation has invalid input", __func__); |
| 752 | } |
| 753 | |
| 754 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 755 | if (!output) |
| 756 | { |
| 757 | return Fail("%s: Could not read output 0", __func__); |
| 758 | } |
| 759 | |
| 760 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 761 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 762 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 763 | ElementwiseUnaryDescriptor descriptor(unaryOperation); |
| 764 | |
| 765 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 766 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 767 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 768 | { |
| 769 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 770 | IsElementwiseUnarySupported, |
| 771 | data.m_Backends, |
| 772 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 773 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 774 | inputInfo, |
| 775 | outputInfo, |
| 776 | descriptor); |
| 777 | }; |
| 778 | |
| 779 | if(!IsDynamicTensor(outputInfo)) |
| 780 | { |
| 781 | validateFunc(outputInfo, isSupported); |
| 782 | } |
| 783 | else |
| 784 | { |
| 785 | isSupported = AreDynamicTensorsSupported(); |
| 786 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 787 | |
| 788 | if (!isSupported) |
| 789 | { |
| 790 | return false; |
| 791 | } |
| 792 | |
| 793 | IConnectableLayer* layer = data.m_Network->AddElementwiseUnaryLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 794 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 795 | if (!layer) |
| 796 | { |
| 797 | return Fail("%s: Could not add the ElementwiseUnaryLayer", __func__); |
| 798 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 799 | input.Connect(layer->GetInputSlot(0)); |
| 800 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 801 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 802 | } |
| 803 | |
| 804 | template<typename HalPolicy, |
| 805 | typename HalOperation = typename HalPolicy::Operation, |
| 806 | typename HalModel = typename HalPolicy::Model> |
| 807 | bool ConvertExpandDims(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 808 | { |
| 809 | using HalOperand = typename HalPolicy::Operand; |
| 810 | using HalOperandType = typename HalPolicy::OperandType; |
| 811 | |
| 812 | ALOGV("HalPolicy::ConvertExpandDims()"); |
| 813 | |
| 814 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 815 | |
| 816 | if (!input.IsValid()) |
| 817 | { |
| 818 | return Fail("%s: Operation has invalid input", __func__); |
| 819 | } |
| 820 | |
| 821 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 822 | if (!output) |
| 823 | { |
| 824 | return Fail("%s: Operation has invalid output", __func__); |
| 825 | } |
| 826 | |
| 827 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 828 | |
| 829 | int32_t axis; |
| 830 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, axis, model, data)) |
| 831 | { |
| 832 | return Fail("%s: failed to get axis input value", __func__); |
| 833 | } |
| 834 | |
| 835 | TensorShape targetShape; |
| 836 | |
| 837 | try |
| 838 | { |
| 839 | targetShape = armnnUtils::ExpandDims(input.GetTensorInfo().GetShape(), axis); |
| 840 | } |
| 841 | catch (const std::exception& e) |
| 842 | { |
| 843 | return Fail("%s: %s", __func__, e.what()); |
| 844 | } |
| 845 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 846 | ReshapeDescriptor reshapeDescriptor; |
| 847 | reshapeDescriptor.m_TargetShape = targetShape; |
| 848 | |
| 849 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 850 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 851 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 852 | { |
| 853 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 854 | IsReshapeSupported, |
| 855 | data.m_Backends, |
| 856 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 857 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 858 | input.GetTensorInfo(), |
| 859 | outputInfo, |
| 860 | reshapeDescriptor); |
| 861 | }; |
| 862 | |
| 863 | if(!IsDynamicTensor(outputInfo)) |
| 864 | { |
Nikhil Raj | c5e0bb0 | 2021-04-02 10:02:16 +0100 | [diff] [blame] | 865 | if (targetShape != outputInfo.GetShape()) |
| 866 | { |
| 867 | return Fail("%s: Shape of the output operand does not match the resolved expanded shape", __func__); |
| 868 | } |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 869 | validateFunc(outputInfo, isSupported); |
| 870 | } |
| 871 | else |
| 872 | { |
| 873 | isSupported = AreDynamicTensorsSupported(); |
| 874 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 875 | |
| 876 | if (!isSupported) |
| 877 | { |
| 878 | return false; |
| 879 | } |
| 880 | |
| 881 | IConnectableLayer* layer = data.m_Network->AddReshapeLayer(reshapeDescriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 882 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 883 | if (!layer) |
| 884 | { |
| 885 | return Fail("%s: Could not add the ReshapeLayer", __func__); |
| 886 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 887 | input.Connect(layer->GetInputSlot(0)); |
| 888 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 889 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 890 | } |
| 891 | |
| 892 | template<typename HalPolicy, |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 893 | typename HalOperation = typename HalPolicy::Operation, |
| 894 | typename HalModel = typename HalPolicy::Model> |
| 895 | bool ConvertGather(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 896 | { |
| 897 | using HalOperand = typename HalPolicy::Operand; |
| 898 | using HalOperandType = typename HalPolicy::OperandType; |
| 899 | |
| 900 | ALOGV("HalPolicy::ConvertGather()"); |
| 901 | |
| 902 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 903 | if (!input.IsValid()) |
| 904 | { |
| 905 | return Fail("%s: Operation has invalid input", __func__); |
| 906 | } |
| 907 | auto inputDimensions = input.GetTensorInfo().GetNumDimensions(); |
| 908 | |
| 909 | LayerInputHandle indices = ConvertToLayerInputHandle<HalPolicy>(operation, 2, model, data); |
| 910 | if (!indices.IsValid()) |
| 911 | { |
| 912 | return Fail("%s: Operation has invalid indices", __func__); |
| 913 | } |
| 914 | auto indicesDimensions = indices.GetTensorInfo().GetNumDimensions(); |
| 915 | |
| 916 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 917 | if (!output) |
| 918 | { |
| 919 | return Fail("%s: Operation has invalid output", __func__); |
| 920 | } |
| 921 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 922 | auto outputDimensions = outputInfo.GetNumDimensions(); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 923 | if (outputDimensions != inputDimensions + indicesDimensions - 1) |
| 924 | { |
| 925 | 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] | 926 | __func__, outputDimensions, inputDimensions, indicesDimensions); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 927 | } |
| 928 | |
Teresa Charlin | e0fd650 | 2022-12-05 16:45:50 +0000 | [diff] [blame] | 929 | int32_t axis; |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 930 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, axis, model, data)) |
| 931 | { |
| 932 | return Fail("%s: Operation has invalid or unsupported axis operand", __func__); |
| 933 | } |
Teresa Charlin | e0fd650 | 2022-12-05 16:45:50 +0000 | [diff] [blame] | 934 | int32_t inputDimensions_int = static_cast<int32_t>(inputDimensions); |
| 935 | if ((axis < -inputDimensions_int) || (inputDimensions_int <= axis)) |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 936 | { |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 937 | return Fail("%s: Operation has invalid axis: %d. It is out of bounds [-%d, %d))", __func__, axis, |
| 938 | inputDimensions, inputDimensions); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 939 | } |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 940 | |
| 941 | GatherDescriptor desc; |
| 942 | desc.m_Axis = axis; |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 943 | |
| 944 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 945 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 946 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 947 | { |
| 948 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 949 | IsGatherSupported, |
| 950 | data.m_Backends, |
| 951 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 952 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 953 | input.GetTensorInfo(), |
| 954 | indices.GetTensorInfo(), |
| 955 | outputInfo, |
| 956 | desc); |
| 957 | }; |
| 958 | |
| 959 | if(!IsDynamicTensor(outputInfo)) |
| 960 | { |
| 961 | validateFunc(outputInfo, isSupported); |
| 962 | } |
| 963 | else |
| 964 | { |
| 965 | isSupported = AreDynamicTensorsSupported(); |
| 966 | } |
| 967 | |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 968 | if (!isSupported) |
| 969 | { |
| 970 | return false; |
| 971 | } |
| 972 | |
Teresa Charlin | 5d4873f | 2020-06-03 14:39:29 +0100 | [diff] [blame] | 973 | IConnectableLayer* layer = data.m_Network->AddGatherLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 974 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 975 | if (!layer) |
| 976 | { |
| 977 | return Fail("%s: Could not add the GatherLayer", __func__); |
| 978 | } |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 979 | input.Connect(layer->GetInputSlot(0)); |
| 980 | indices.Connect(layer->GetInputSlot(1)); |
| 981 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 982 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Teresa Charlin | f931af9 | 2020-04-10 16:46:53 +0100 | [diff] [blame] | 983 | } |
| 984 | |
| 985 | template<typename HalPolicy, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 986 | typename HalOperation = typename HalPolicy::Operation, |
| 987 | typename HalModel = typename HalPolicy::Model> |
| 988 | bool ConvertGroupedConv2d(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 989 | { |
| 990 | using HalOperand = typename HalPolicy::Operand; |
| 991 | using HalOperandType = typename HalPolicy::OperandType; |
| 992 | |
| 993 | ALOGV("HalPolicy::ConvertGroupedConv2d()"); |
| 994 | |
| 995 | // |
| 996 | // Parse data |
| 997 | // |
| 998 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 999 | if (!input.IsValid()) |
| 1000 | { |
| 1001 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1002 | } |
| 1003 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1004 | |
| 1005 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1006 | if (!output) |
| 1007 | { |
| 1008 | return Fail("%s: Could not read output 0", __func__); |
| 1009 | } |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1010 | TensorInfo outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1011 | |
| 1012 | // Look ahead to determine data layout |
| 1013 | DataLayout dataLayout = DataLayout::NHWC; |
| 1014 | if (operation.inputs.size() == 12) |
| 1015 | { |
| 1016 | dataLayout = OptionalDataLayout<HalPolicy>(operation, 11, model, data); |
| 1017 | } |
| 1018 | else |
| 1019 | { |
| 1020 | dataLayout = OptionalDataLayout<HalPolicy>(operation, 8, model, data); |
| 1021 | } |
| 1022 | |
| 1023 | // NOTE: |
| 1024 | // NNAPI weights are always OHWI, i.e. [depth_out, filter_height, filter_width, depth_group], |
| 1025 | // but Arm NN expects the filter's height and width indices to match the input's height and |
| 1026 | // width indices so when the DataLayout is NCHW, we need to permute the weights to OIHW |
| 1027 | const PermutationVector ohwiToOihw = { 0u, 2u, 3u, 1u }; |
| 1028 | const ConstTensorPin weightsPin = (dataLayout == DataLayout::NCHW) ? |
| 1029 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, |
| 1030 | model, data, ohwiToOihw) : |
| 1031 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 1032 | const ConstTensorPin biasesPin = |
| 1033 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 1034 | if (!weightsPin.IsValid() || !biasesPin.IsValid()) |
| 1035 | { |
| 1036 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1037 | } |
| 1038 | |
| 1039 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 1040 | ConstTensor biases = biasesPin.GetConstTensor(); |
| 1041 | SanitizeBiasQuantizationScale(biases.GetInfo(), weights.GetInfo(), inputInfo); |
| 1042 | |
| 1043 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 1044 | const TensorShape& outputShape = outputInfo.GetShape(); |
| 1045 | const TensorShape& weightsShape = weights.GetShape(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1046 | |
| 1047 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| 1048 | const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| 1049 | const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 1050 | const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 1051 | |
| 1052 | Convolution2dDescriptor desc; |
| 1053 | desc.m_DataLayout = dataLayout; |
| 1054 | desc.m_BiasEnabled = true; |
| 1055 | |
Finn Williams | f769f29 | 2021-06-25 12:53:09 +0100 | [diff] [blame] | 1056 | unsigned int numGroups; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1057 | ActivationFn activation; |
| 1058 | |
| 1059 | if (operation.inputs.size() == 12) |
| 1060 | { |
| 1061 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 1062 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 1063 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 1064 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 1065 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 1066 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 1067 | !GetInputScalar<HalPolicy>(operation, 9, HalOperandType::INT32, numGroups, model, data) || |
| 1068 | !GetInputActivationFunction<HalPolicy>(operation, 10, activation, model, data)) |
| 1069 | { |
| 1070 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 1071 | } |
| 1072 | |
| 1073 | } |
| 1074 | else if (operation.inputs.size() == 9) |
| 1075 | { |
| 1076 | android::nn::PaddingScheme paddingScheme; |
| 1077 | if (!GetInputPaddingScheme<HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 1078 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 1079 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 1080 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, numGroups, model, data) || |
| 1081 | !GetInputActivationFunction<HalPolicy>(operation, 7, activation, model, data)) |
| 1082 | { |
| 1083 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 1084 | } |
| 1085 | |
| 1086 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 1087 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 1088 | |
| 1089 | const uint32_t kernelX = weightsShape[widthIndex]; |
| 1090 | const uint32_t kernelY = weightsShape[heightIndex]; |
| 1091 | |
| 1092 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 1093 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 1094 | } |
| 1095 | else |
| 1096 | { |
| 1097 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 1098 | } |
| 1099 | |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1100 | // Equivalent to outputShape[channelsIndex], but we can't know the outputShape in the case of dynamic tensors |
| 1101 | const unsigned int outputChannels = weightsShape[0]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1102 | |
| 1103 | const unsigned int channelsPerGroup = weightsShape[channelsIndex]; |
| 1104 | const unsigned int channelMultiplier = outputChannels / numGroups; |
| 1105 | |
| 1106 | // |
| 1107 | // Validate all relevant inputs |
| 1108 | // |
| 1109 | if (numGroups <= 0) |
| 1110 | { |
| 1111 | return Fail("%s: Number of groups must be greater than 0. Got: %d", __func__, numGroups); |
| 1112 | } |
| 1113 | |
| 1114 | if (outputChannels % numGroups != 0u) |
| 1115 | { |
| 1116 | return Fail("%s: Output channels must be divisible by the number of groups", __func__); |
| 1117 | } |
| 1118 | |
| 1119 | // |
| 1120 | // Set up Splitter layer |
| 1121 | // |
| 1122 | unsigned int splitterDimSizes[4] = { inputShape[0], inputShape[1], inputShape[2], inputShape[3] }; |
| 1123 | splitterDimSizes[channelsIndex] /= numGroups; // split in depth |
| 1124 | |
| 1125 | TensorInfo splitterOutputInfo(4, |
| 1126 | splitterDimSizes, |
| 1127 | inputInfo.GetDataType(), |
| 1128 | inputInfo.GetQuantizationScale(), |
| 1129 | inputInfo.GetQuantizationOffset()); |
| 1130 | |
| 1131 | std::vector<std::reference_wrapper<TensorInfo>> splitterOutputInfos(numGroups, std::ref(splitterOutputInfo)); |
| 1132 | |
| 1133 | ViewsDescriptor splitterDesc(numGroups); |
Mike Kelly | 23eccd8 | 2023-08-08 12:29:14 +0100 | [diff] [blame] | 1134 | splitterDesc.SetAxis(armnn::numeric_cast<int32_t>(channelsIndex)); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1135 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1136 | { |
| 1137 | splitterDesc.SetViewOriginCoord(group, channelsIndex, splitterDimSizes[channelsIndex] * group); |
| 1138 | for (unsigned int dimIdx = 0u; dimIdx < 4u; dimIdx++) |
| 1139 | { |
| 1140 | splitterDesc.SetViewSize(group, dimIdx, splitterDimSizes[dimIdx]); |
| 1141 | } |
| 1142 | } |
| 1143 | |
| 1144 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1145 | armnn::BackendId setBackendSplit; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1146 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1147 | IsSplitterSupported, |
| 1148 | data.m_Backends, |
| 1149 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1150 | setBackendSplit, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1151 | inputInfo, |
| 1152 | splitterOutputInfos, |
| 1153 | splitterDesc); |
| 1154 | if (!isSupported) |
| 1155 | { |
| 1156 | return false; |
| 1157 | } |
| 1158 | |
| 1159 | IConnectableLayer* splitterLayer = data.m_Network->AddSplitterLayer(splitterDesc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1160 | splitterLayer->SetBackendId(setBackendSplit); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1161 | if (!splitterLayer) |
| 1162 | { |
| 1163 | return Fail("%s: Failed to add SplitterLayer", __func__); |
| 1164 | } |
| 1165 | |
| 1166 | input.Connect(splitterLayer->GetInputSlot(0)); |
| 1167 | for (unsigned int group = 0u; group < splitterLayer->GetNumOutputSlots(); ++group) |
| 1168 | { |
| 1169 | splitterLayer->GetOutputSlot(group).SetTensorInfo(splitterOutputInfo); |
| 1170 | } |
| 1171 | |
| 1172 | // |
| 1173 | // Set up Convolution2d layers for each group |
| 1174 | // |
| 1175 | |
| 1176 | // Set up group tensor shapes |
| 1177 | TensorShape groupInputShape(inputShape); |
| 1178 | groupInputShape[channelsIndex] = channelsPerGroup; |
| 1179 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1180 | TensorShape groupWeightsShape(weightsShape); |
| 1181 | groupWeightsShape[0] /= channelMultiplier * numGroups; |
| 1182 | |
| 1183 | TensorShape groupBiasesShape({ 1 }); |
| 1184 | |
| 1185 | // Set up group tensor infos |
| 1186 | TensorInfo groupInputInfo(inputInfo); |
| 1187 | groupInputInfo.SetShape(groupInputShape); |
| 1188 | |
| 1189 | const TensorInfo& weightsInfo = weights.GetInfo(); |
| 1190 | TensorInfo groupWeightsInfo(weightsInfo); |
| 1191 | groupWeightsInfo.SetShape(groupWeightsShape); |
| 1192 | |
| 1193 | const TensorInfo& biasesInfo = biases.GetInfo(); |
| 1194 | TensorInfo groupBiasesInfo(biasesInfo); |
| 1195 | groupBiasesInfo.SetShape(groupBiasesShape); |
| 1196 | |
| 1197 | TensorInfo groupOutputInfo(outputInfo); |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1198 | |
| 1199 | TensorShape groupOutputShape(outputShape); |
| 1200 | const bool isDynamic = IsDynamicTensor(outputInfo); |
| 1201 | if (!isDynamic) |
| 1202 | { |
| 1203 | groupOutputShape[channelsIndex] = 1; |
| 1204 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1205 | groupOutputInfo.SetShape(groupOutputShape); |
| 1206 | |
| 1207 | const unsigned int weightsDataTypeSize = GetDataTypeSize(groupWeightsInfo.GetDataType()); |
| 1208 | const unsigned int biasesDataTypeSize = GetDataTypeSize(groupBiasesInfo.GetDataType()); |
| 1209 | |
| 1210 | std::vector<IConnectableLayer*> convLayers(numGroups * channelMultiplier, nullptr); |
| 1211 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1212 | { |
| 1213 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1214 | { |
| 1215 | auto index = group * channelMultiplier + m; |
| 1216 | |
| 1217 | const unsigned int weightsDataOffset = groupWeightsShape.GetNumElements() * index * weightsDataTypeSize; |
| 1218 | const unsigned int biasesDataOffset = groupBiasesShape.GetNumElements() * index * biasesDataTypeSize; |
| 1219 | |
| 1220 | if (weightsInfo.HasPerAxisQuantization()) |
| 1221 | { |
| 1222 | // Extract per-axis quantization scales for group weights |
| 1223 | const std::vector<float>& weightsQuantScales = weightsInfo.GetQuantizationScales(); |
| 1224 | groupWeightsInfo.SetQuantizationScales( |
| 1225 | std::vector<float>(weightsQuantScales.begin() + index, |
| 1226 | weightsQuantScales.begin() + index + groupWeightsShape[0])); |
| 1227 | |
| 1228 | // Extract per-axis quantization scales for group biases |
| 1229 | const std::vector<float>& biasesQuantScales = biasesInfo.GetQuantizationScales(); |
| 1230 | groupBiasesInfo.SetQuantizationScales( |
| 1231 | std::vector<float>(biasesQuantScales.begin() + index, |
| 1232 | biasesQuantScales.begin() + index + groupWeightsShape[0])); |
| 1233 | } |
| 1234 | |
| 1235 | // Extract weights and biases data for current group convolution |
| 1236 | ConstTensor groupWeights(groupWeightsInfo, |
| 1237 | static_cast<const void *>(reinterpret_cast<const char *>(weights.GetMemoryArea()) + |
| 1238 | weightsDataOffset)); |
| 1239 | ConstTensor groupBiases(groupBiasesInfo, |
| 1240 | static_cast<const void *>(reinterpret_cast<const char *>(biases.GetMemoryArea()) + |
| 1241 | biasesDataOffset)); |
| 1242 | |
| 1243 | isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1244 | armnn::BackendId setBackendConv; |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1245 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1246 | { |
| 1247 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1248 | IsConvolution2dSupported, |
| 1249 | data.m_Backends, |
| 1250 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1251 | setBackendConv, |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1252 | groupInputInfo, |
| 1253 | outputInfo, |
| 1254 | desc, |
| 1255 | groupWeightsInfo, |
| 1256 | Optional<TensorInfo>(groupBiasesInfo)); |
| 1257 | }; |
| 1258 | |
| 1259 | if(!isDynamic) |
| 1260 | { |
| 1261 | validateFunc(groupOutputInfo, isSupported); |
| 1262 | } |
| 1263 | else |
| 1264 | { |
| 1265 | isSupported = AreDynamicTensorsSupported(); |
| 1266 | } |
| 1267 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1268 | if (!isSupported) |
| 1269 | { |
| 1270 | return false; |
| 1271 | } |
| 1272 | |
Teresa Charlin | f869626 | 2022-08-30 15:55:28 +0100 | [diff] [blame] | 1273 | IConnectableLayer* weightsLayer = data.m_Network->AddConstantLayer(groupWeights); |
| 1274 | IConnectableLayer* biasLayer = data.m_Network->AddConstantLayer(groupBiases); |
| 1275 | IConnectableLayer* convLayer = data.m_Network->AddConvolution2dLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1276 | convLayer->SetBackendId(setBackendConv); |
Keith Davis | 8f22bed | 2022-04-29 10:57:27 +0100 | [diff] [blame] | 1277 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1278 | if (!convLayer) |
| 1279 | { |
| 1280 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 1281 | } |
| 1282 | |
| 1283 | splitterLayer->GetOutputSlot(group).Connect(convLayer->GetInputSlot(0)); |
Teresa Charlin | f869626 | 2022-08-30 15:55:28 +0100 | [diff] [blame] | 1284 | weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1)); |
| 1285 | biasLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(2)); |
| 1286 | |
| 1287 | weightsLayer->GetOutputSlot(0).SetTensorInfo(groupWeightsInfo); |
| 1288 | biasLayer->GetOutputSlot(0).SetTensorInfo(groupBiasesInfo); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1289 | convLayer->GetOutputSlot(0).SetTensorInfo(groupOutputInfo); |
| 1290 | |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1291 | if(isDynamic) |
| 1292 | { |
| 1293 | convLayer->GetOutputSlot(0).IsTensorInfoSet(); |
| 1294 | |
| 1295 | validateFunc(convLayer->GetOutputSlot(0).GetTensorInfo(), isSupported); |
| 1296 | |
| 1297 | outputInfo = convLayer->GetOutputSlot(0).GetTensorInfo(); |
| 1298 | |
| 1299 | if (!isSupported) |
| 1300 | { |
| 1301 | return false; |
| 1302 | } |
| 1303 | } |
| 1304 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1305 | convLayers[index] = convLayer; |
| 1306 | } |
| 1307 | } |
| 1308 | |
| 1309 | // |
| 1310 | // Set up Concat layer |
| 1311 | // |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1312 | ConcatDescriptor concatDescriptor; |
| 1313 | // Equivalent to outputShape[channelsIndex], but we can't know the outputShape in the case of dynamic tensors |
| 1314 | concatDescriptor = ConcatDescriptor(weightsShape[0]); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1315 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1316 | { |
| 1317 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1318 | { |
| 1319 | auto index = group * channelMultiplier + m; |
| 1320 | concatDescriptor.SetViewOriginCoord(index, channelsIndex, index); |
| 1321 | concatDescriptor.SetConcatAxis(channelsIndex); |
| 1322 | } |
| 1323 | } |
| 1324 | |
| 1325 | isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1326 | armnn::BackendId setBackendConcat; |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1327 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1328 | IsConcatSupported, |
| 1329 | data.m_Backends, |
| 1330 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1331 | setBackendConcat, |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1332 | std::vector<const TensorInfo*>(numGroups * channelMultiplier, &groupOutputInfo), |
| 1333 | outputInfo, |
| 1334 | concatDescriptor); |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1335 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1336 | if (!isSupported) |
| 1337 | { |
| 1338 | return false; |
| 1339 | } |
| 1340 | |
| 1341 | IConnectableLayer* concatLayer = data.m_Network->AddConcatLayer(concatDescriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1342 | concatLayer->SetBackendId(setBackendConcat); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1343 | if (!concatLayer) |
| 1344 | { |
| 1345 | return Fail("%s: AddConcatLayer failed", __func__); |
| 1346 | } |
| 1347 | |
| 1348 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1349 | { |
| 1350 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1351 | { |
| 1352 | auto index = group * channelMultiplier + m; |
| 1353 | convLayers[index]->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(index)); |
| 1354 | } |
| 1355 | } |
| 1356 | concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1357 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 1358 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *concatLayer, model, |
Finn Williams | b033117 | 2020-10-08 14:33:13 +0100 | [diff] [blame] | 1359 | data, nullptr, nullptr, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1360 | } |
| 1361 | |
| 1362 | template<typename HalPolicy, |
| 1363 | typename HalOperation = typename HalPolicy::Operation, |
| 1364 | typename HalModel = typename HalPolicy::Model> |
| 1365 | bool ConvertInstanceNormalization(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1366 | { |
| 1367 | using HalOperand = typename HalPolicy::Operand; |
| 1368 | using HalOperandType = typename HalPolicy::OperandType; |
| 1369 | |
| 1370 | ALOGV("HalPolicy::ConvertInstanceNormalization()"); |
| 1371 | |
| 1372 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1373 | if (!input.IsValid()) |
| 1374 | { |
| 1375 | return Fail("%s: Operation has an invalid input 0", __func__); |
| 1376 | } |
| 1377 | |
| 1378 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1379 | if (!output) |
| 1380 | { |
| 1381 | return Fail("%s: Operation has an invalid output", __func__); |
| 1382 | } |
| 1383 | |
| 1384 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1385 | |
| 1386 | // Determine data type of input tensor |
| 1387 | HalOperandType inputType; |
| 1388 | if (!GetOperandType<HalPolicy>(operation, 0, model, inputType)) |
| 1389 | { |
| 1390 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1391 | } |
| 1392 | |
| 1393 | InstanceNormalizationDescriptor desc; |
| 1394 | |
| 1395 | // Read gamma, beta & epsilon |
| 1396 | if (inputType == HalOperandType::TENSOR_FLOAT16) |
| 1397 | { |
| 1398 | Half fp16Gamma; |
| 1399 | Half fp16Beta; |
| 1400 | Half fp16Epsilon; |
| 1401 | |
| 1402 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, fp16Gamma, model, data) || |
| 1403 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::FLOAT16, fp16Beta, model, data) || |
| 1404 | !GetInputScalar<HalPolicy>(operation, 3, HalOperandType::FLOAT16, fp16Epsilon, model, data)) |
| 1405 | { |
| 1406 | return Fail("%s: Operation has invalid inputs (FLOAT16)", __func__); |
| 1407 | } |
| 1408 | |
| 1409 | desc.m_Gamma = static_cast<float>(fp16Gamma); |
| 1410 | desc.m_Beta = static_cast<float>(fp16Beta); |
| 1411 | desc.m_Eps = static_cast<float>(fp16Epsilon); |
| 1412 | } |
| 1413 | else if (inputType == HalOperandType::TENSOR_FLOAT32) |
| 1414 | { |
| 1415 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT32, desc.m_Gamma, model, data) || |
| 1416 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::FLOAT32, desc.m_Beta, model, data) || |
| 1417 | !GetInputScalar<HalPolicy>(operation, 3, HalOperandType::FLOAT32, desc.m_Eps, model, data)) |
| 1418 | { |
| 1419 | return Fail("%s: Operation has invalid inputs (FLOAT32)", __func__); |
| 1420 | } |
| 1421 | } |
| 1422 | else |
| 1423 | { |
| 1424 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1425 | } |
| 1426 | |
| 1427 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 4, model, data); |
| 1428 | |
| 1429 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1430 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1431 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1432 | { |
| 1433 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1434 | IsInstanceNormalizationSupported, |
| 1435 | data.m_Backends, |
| 1436 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1437 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1438 | input.GetTensorInfo(), |
| 1439 | outputInfo, |
| 1440 | desc); |
| 1441 | }; |
| 1442 | |
| 1443 | if(IsDynamicTensor(outputInfo)) |
| 1444 | { |
| 1445 | isSupported = AreDynamicTensorsSupported(); |
| 1446 | } |
| 1447 | else |
| 1448 | { |
| 1449 | validateFunc(outputInfo, isSupported); |
| 1450 | } |
| 1451 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1452 | if (!isSupported) |
| 1453 | { |
| 1454 | return false; |
| 1455 | } |
| 1456 | |
| 1457 | IConnectableLayer* layer = data.m_Network->AddInstanceNormalizationLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1458 | layer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1459 | input.Connect(layer->GetInputSlot(0)); |
| 1460 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1461 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1462 | } |
| 1463 | |
| 1464 | template<typename HalPolicy, |
| 1465 | typename HalOperation = typename HalPolicy::Operation, |
| 1466 | typename HalModel = typename HalPolicy::Model> |
| 1467 | bool ConvertLogSoftmax(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1468 | { |
| 1469 | using HalOperand = typename HalPolicy::Operand; |
| 1470 | using HalOperandType = typename HalPolicy::OperandType; |
| 1471 | |
| 1472 | ALOGV("HalPolicy::ConvertLogSoftmax()"); |
| 1473 | |
| 1474 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1475 | if (!input.IsValid()) |
| 1476 | { |
| 1477 | return Fail("%s: Failed to read input 0", __func__); |
| 1478 | } |
| 1479 | |
| 1480 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1481 | if (!output) |
| 1482 | { |
| 1483 | return Fail("%s: Failed to read output", __func__); |
| 1484 | } |
| 1485 | |
| 1486 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1487 | |
| 1488 | // Determine data type of input tensor |
| 1489 | HalOperandType inputType; |
| 1490 | if (!GetOperandType<HalPolicy>(operation, 0, model, inputType)) |
| 1491 | { |
| 1492 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1493 | } |
| 1494 | |
| 1495 | LogSoftmaxDescriptor descriptor; |
| 1496 | |
| 1497 | // Read beta |
| 1498 | if (inputType == HalOperandType::TENSOR_FLOAT16) |
| 1499 | { |
| 1500 | Half fp16Beta; |
| 1501 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, fp16Beta, model, data)) |
| 1502 | { |
| 1503 | return Fail("%s: Failed to read input 1 (FLOAT16)", __func__); |
| 1504 | } |
| 1505 | |
| 1506 | descriptor.m_Beta = static_cast<float>(fp16Beta); |
| 1507 | } |
| 1508 | else if (inputType == HalOperandType::TENSOR_FLOAT32) |
| 1509 | { |
| 1510 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT32, descriptor.m_Beta, model, data)) |
| 1511 | { |
| 1512 | return Fail("%s: Failed to read input 1 (FLOAT32)", __func__); |
| 1513 | } |
| 1514 | } |
| 1515 | else |
| 1516 | { |
| 1517 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1518 | } |
| 1519 | |
| 1520 | // Read axis |
| 1521 | if (!GetInputInt32<HalPolicy>(operation, 2, descriptor.m_Axis, model, data)) |
| 1522 | { |
| 1523 | return Fail("%s: Failed to read input 2", __func__); |
| 1524 | } |
| 1525 | |
| 1526 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1527 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1528 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1529 | { |
| 1530 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1531 | IsLogSoftmaxSupported, |
| 1532 | data.m_Backends, |
| 1533 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1534 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1535 | input.GetTensorInfo(), |
| 1536 | outputInfo, |
| 1537 | descriptor); |
| 1538 | }; |
| 1539 | |
| 1540 | if(IsDynamicTensor(outputInfo)) |
| 1541 | { |
| 1542 | isSupported = AreDynamicTensorsSupported(); |
| 1543 | } |
| 1544 | else |
| 1545 | { |
| 1546 | validateFunc(outputInfo, isSupported); |
| 1547 | } |
| 1548 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1549 | if (!isSupported) |
| 1550 | { |
| 1551 | return false; |
| 1552 | } |
| 1553 | |
| 1554 | IConnectableLayer* layer = data.m_Network->AddLogSoftmaxLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1555 | layer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1556 | if (!layer) |
| 1557 | { |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1558 | return Fail("%s: Could not add the LogSoftmaxLayer", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1559 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1560 | input.Connect(layer->GetInputSlot(0)); |
| 1561 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1562 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1563 | } |
| 1564 | |
| 1565 | template<typename HalPolicy, |
| 1566 | typename HalOperation = typename HalPolicy::Operation, |
| 1567 | typename HalModel = typename HalPolicy::Model> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1568 | bool ConvertPadV2(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1569 | { |
| 1570 | using HalOperand = typename HalPolicy::Operand; |
| 1571 | using HalOperandType = typename HalPolicy::OperandType; |
| 1572 | |
| 1573 | ALOGV("HalPolicy::ConvertPadV2()"); |
| 1574 | |
| 1575 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1576 | if (!input.IsValid()) |
| 1577 | { |
| 1578 | return Fail("%s: Could not read input 0", __func__); |
| 1579 | } |
| 1580 | |
| 1581 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1582 | if (!output) |
| 1583 | { |
| 1584 | return Fail("%s: Could not read output", __func__); |
| 1585 | } |
| 1586 | |
| 1587 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1588 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 1589 | |
| 1590 | PadDescriptor descriptor; |
| 1591 | if (!ConvertPaddings<HalPolicy>(operation, model, data, rank, descriptor)) |
| 1592 | { |
| 1593 | return Fail("%s: Could not convert paddings", __func__); |
| 1594 | } |
| 1595 | |
| 1596 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1597 | |
| 1598 | // Determine type of padding value |
| 1599 | HalOperandType operandType0; |
| 1600 | HalOperandType operandType2; |
| 1601 | |
| 1602 | if (!GetOperandType<HalPolicy>(operation, 0, model, operandType0) || |
| 1603 | !GetOperandType<HalPolicy>(operation, 2, model, operandType2)) |
| 1604 | { |
| 1605 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1606 | } |
| 1607 | |
| 1608 | // Read value to use for padding |
| 1609 | if (operandType0 == HalOperandType::TENSOR_FLOAT16 && operandType2 == HalOperandType::FLOAT16) |
| 1610 | { |
| 1611 | Half f16PadValue; |
| 1612 | if (!GetInputScalar<HalPolicy>(operation, 2, operandType2, f16PadValue, model, data)) |
| 1613 | { |
| 1614 | return Fail("%s: Could not read input 2 (FLOAT16)", __func__); |
| 1615 | } |
| 1616 | |
| 1617 | descriptor.m_PadValue = f16PadValue; |
| 1618 | } |
| 1619 | else if (operandType0 == HalOperandType::TENSOR_FLOAT32 && operandType2 == HalOperandType::FLOAT32) |
| 1620 | { |
| 1621 | if (!GetInputFloat32<HalPolicy>(operation, 2, descriptor.m_PadValue, model, data)) |
| 1622 | { |
| 1623 | return Fail("%s: Could not read input 2 (FLOAT32)", __func__); |
| 1624 | } |
| 1625 | } |
Teresa Charlin | d3381d5 | 2021-06-02 18:35:16 +0100 | [diff] [blame] | 1626 | else if (isQuantizedOperand(operandType0) && operandType2 == HalOperandType::INT32) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1627 | { |
| 1628 | int32_t intPadValue = 0; |
| 1629 | if (!GetInputInt32<HalPolicy>(operation, 2, intPadValue, model, data)) |
| 1630 | { |
| 1631 | return Fail("%s: Could not read input 2 (INT32)", __func__); |
| 1632 | } |
| 1633 | descriptor.m_PadValue = intPadValue; |
| 1634 | } |
| 1635 | else |
| 1636 | { |
| 1637 | return Fail("%s: Operation has invalid inputs: type mismatch", __func__); |
| 1638 | } |
| 1639 | |
| 1640 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1641 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1642 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1643 | { |
| 1644 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1645 | IsPadSupported, |
| 1646 | data.m_Backends, |
| 1647 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1648 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1649 | inputInfo, |
| 1650 | outputInfo, |
| 1651 | descriptor); |
| 1652 | }; |
| 1653 | |
| 1654 | if(IsDynamicTensor(outputInfo)) |
| 1655 | { |
| 1656 | isSupported = AreDynamicTensorsSupported(); |
| 1657 | } |
| 1658 | else |
| 1659 | { |
| 1660 | validateFunc(outputInfo, isSupported); |
| 1661 | } |
| 1662 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1663 | if (!isSupported) |
| 1664 | { |
| 1665 | return false; |
| 1666 | } |
| 1667 | |
| 1668 | IConnectableLayer* const layer = data.m_Network->AddPadLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1669 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1670 | if (!layer) |
| 1671 | { |
| 1672 | return Fail("%s: Could not add the PadLayer", __func__); |
| 1673 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1674 | input.Connect(layer->GetInputSlot(0)); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1675 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1676 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1677 | } |
| 1678 | |
| 1679 | template<typename HalPolicy, |
| 1680 | typename HalOperation = typename HalPolicy::Operation, |
| 1681 | typename HalModel = typename HalPolicy::Model> |
| 1682 | bool ConvertPrelu(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1683 | { |
| 1684 | using HalOperand = typename HalPolicy::Operand; |
| 1685 | |
| 1686 | ALOGV("HalPolicy::ConvertPrelu()"); |
| 1687 | |
| 1688 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1689 | LayerInputHandle alpha = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data); |
| 1690 | |
| 1691 | if (!input.IsValid() || !alpha.IsValid()) |
| 1692 | { |
| 1693 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1694 | } |
| 1695 | |
| 1696 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1697 | |
| 1698 | if (!output) |
| 1699 | { |
| 1700 | return Fail("%s: Could not read output", __func__); |
| 1701 | } |
| 1702 | |
| 1703 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1704 | const TensorInfo& alphaInfo = alpha.GetTensorInfo(); |
| 1705 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1706 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1707 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1708 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1709 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1710 | { |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1711 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1712 | IsPreluSupported, |
| 1713 | data.m_Backends, |
| 1714 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1715 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1716 | inputInfo, |
| 1717 | alphaInfo, |
| 1718 | outputInfo); |
| 1719 | }; |
| 1720 | |
| 1721 | if(IsDynamicTensor(outputInfo)) |
| 1722 | { |
| 1723 | isSupported = AreDynamicTensorsSupported(); |
| 1724 | } |
| 1725 | else |
| 1726 | { |
| 1727 | validateFunc(outputInfo, isSupported); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1728 | } |
| 1729 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1730 | if (!isSupported) |
| 1731 | { |
| 1732 | return false; |
| 1733 | } |
| 1734 | |
| 1735 | IConnectableLayer* const layer = data.m_Network->AddPreluLayer(); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1736 | layer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1737 | if (!layer) |
| 1738 | { |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1739 | return Fail("%s: Could not add the PreluLayer", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1740 | } |
| 1741 | |
| 1742 | bool isReshapeSupported = BroadcastTensor(input, alpha, layer, data); |
| 1743 | if (!isReshapeSupported) |
| 1744 | { |
| 1745 | return false; |
| 1746 | } |
| 1747 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1748 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1749 | } |
| 1750 | |
| 1751 | template<typename HalPolicy, |
| 1752 | typename HalOperation = typename HalPolicy::Operation, |
| 1753 | typename HalModel = typename HalPolicy::Model> |
| 1754 | bool ConvertQuantize(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 1755 | { |
| 1756 | using HalOperand = typename HalPolicy::Operand; |
| 1757 | |
| 1758 | ALOGV("HalPolicy::ConvertQuantize()"); |
| 1759 | |
| 1760 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1761 | if (!input.IsValid()) |
| 1762 | { |
| 1763 | return Fail("%s: Operation has invalid input", __func__); |
| 1764 | } |
| 1765 | |
| 1766 | const HalOperand* const outputOperand = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1767 | if (!outputOperand) |
| 1768 | { |
| 1769 | return Fail("%s: Operation has invalid outputs", __func__); |
| 1770 | } |
| 1771 | |
| 1772 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1773 | |
| 1774 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1775 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1776 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1777 | { |
| 1778 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1779 | IsQuantizeSupported, |
| 1780 | data.m_Backends, |
| 1781 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1782 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1783 | input.GetTensorInfo(), |
| 1784 | outputInfo); |
| 1785 | }; |
| 1786 | |
| 1787 | if(IsDynamicTensor(outputInfo)) |
| 1788 | { |
| 1789 | isSupported = AreDynamicTensorsSupported(); |
| 1790 | } |
| 1791 | else |
| 1792 | { |
| 1793 | validateFunc(outputInfo, isSupported); |
| 1794 | } |
| 1795 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1796 | if (!isSupported) |
| 1797 | { |
| 1798 | return false; |
| 1799 | } |
| 1800 | |
| 1801 | IConnectableLayer* const layer = data.m_Network->AddQuantizeLayer(); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 1802 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 1803 | if (!layer) |
| 1804 | { |
| 1805 | return Fail("%s: Could not add the QuantizeLayer", __func__); |
| 1806 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1807 | input.Connect(layer->GetInputSlot(0)); |
| 1808 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 1809 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1810 | } |
| 1811 | |
| 1812 | template<typename HalPolicy, |
| 1813 | typename HalOperation = typename HalPolicy::Operation, |
| 1814 | typename HalModel = typename HalPolicy::Model> |
Sadik Armagan | 813f230 | 2020-05-19 14:10:30 +0100 | [diff] [blame] | 1815 | bool ConvertQuantized16BitLstm(const HalOperation& operation, const HalModel& model, ConversionData& data) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1816 | { |
| 1817 | using HalOperand = typename HalPolicy::Operand; |
| 1818 | |
Sadik Armagan | 813f230 | 2020-05-19 14:10:30 +0100 | [diff] [blame] | 1819 | ALOGV("HalPolicy::ConvertQuantized16BitLstm()"); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1820 | |
| 1821 | //Inputs: |
| 1822 | // 0: The input: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBatches, inputSize] |
| 1823 | // specifying the input to the LSTM cell. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 1824 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 1825 | if (!input.IsValid()) |
| 1826 | { |
| 1827 | return Fail("%s: Could not read input 0: input", __func__); |
| 1828 | } |
| 1829 | |
| 1830 | //13: The previous cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape |
| 1831 | // [numBatches, outputSize] specifying the cell state from the previous time step of the LSTM cell. |
| 1832 | // It is quantized using a quantization range of -2^4, 2^4 * 32767/32768. |
| 1833 | LayerInputHandle previousCellStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 13, model, data); |
| 1834 | if (!previousCellStateIn.IsValid()) |
| 1835 | { |
| 1836 | return Fail("%s: Could not read input 13: previousCellStateIn", __func__); |
| 1837 | } |
| 1838 | |
| 1839 | // 14: The previous output state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1840 | // [numBathes, outputSize] specifying the output of the LSTM cell from previous time-step. Tensor |
| 1841 | // is quantized with a fixed quantization range of -1, 127/128. |
| 1842 | LayerInputHandle previousOutputIn = ConvertToLayerInputHandle<HalPolicy>(operation, 14, model, data); |
| 1843 | if (!previousOutputIn.IsValid()) |
| 1844 | { |
| 1845 | return Fail("%s: Could not read input 14: previousOutputIn", __func__); |
| 1846 | } |
| 1847 | |
| 1848 | // Get the input tensors: |
| 1849 | // 1: The input-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1850 | // [outputSize, inputSize] specifying input-to-input part of weights for fully-connected layer inside the |
| 1851 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1852 | const ConstTensorPin inputToInputWeightsPin = |
| 1853 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 1854 | |
| 1855 | // 2: The input-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1856 | // [outputSize, inputSize] specifying input-to-forget part of weights for fully-connected layer inside the |
| 1857 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1858 | const ConstTensorPin inputToForgetWeightsPin = |
| 1859 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 1860 | |
| 1861 | // 3: The input-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1862 | // [outputSize, inputSize] specifying input-to-cell part of weights for fully-connected layer inside the |
| 1863 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1864 | const ConstTensorPin inputToCellWeightsPin = |
| 1865 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 3, model, data); |
| 1866 | |
| 1867 | // 4: The input-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1868 | // [outputSize, inputSize] specifying input-to-output part of weights for fully-connected layer inside the |
| 1869 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1870 | const ConstTensorPin inputToOutputWeightsPin = |
| 1871 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 4, model, data); |
| 1872 | |
| 1873 | // 5: The recurrent-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1874 | // [outputSize, outputSize] specifying recurrent-to-input part of weights for fully-connected layer inside |
| 1875 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1876 | const ConstTensorPin recurrentToInputWeightsPin = |
| 1877 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 5, model, data); |
| 1878 | |
| 1879 | // 6: The recurrent-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1880 | // [outputSize, outputSize] specifying recurrent-to-forget part of weights for fully-connected layer inside |
| 1881 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1882 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 1883 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 6, model, data); |
| 1884 | |
| 1885 | // 7: The recurrent-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1886 | // [outputSize, outputSize] specifying recurrent-to-cell part of weights for fully-connected layer inside |
| 1887 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1888 | const ConstTensorPin recurrentToCellWeightsPin = |
| 1889 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 7, model, data); |
| 1890 | |
| 1891 | // 8: The recurrent-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1892 | // [outputSize, outputSize] specifying recurrent-to-output part of weights for fully-connected layer inside |
| 1893 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1894 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 1895 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 8, model, data); |
| 1896 | |
| 1897 | // 9: The input gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the |
| 1898 | // bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1899 | // of input and weights scales and zeroPoint equal to 0. |
| 1900 | const ConstTensorPin inputGateBiasPin = |
| 1901 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 9, model, data); |
| 1902 | |
| 1903 | // 10: The forget gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 1904 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1905 | // of input and weights scales and zeroPoint equal to 0. |
| 1906 | const ConstTensorPin forgetGateBiasPin = |
| 1907 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 10, model, data); |
| 1908 | |
| 1909 | // 11:The cell bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the bias |
| 1910 | // for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product of input |
| 1911 | // and weights scales and zeroPoint equal to 0. |
| 1912 | const ConstTensorPin cellBiasPin = |
| 1913 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 11, model, data); |
| 1914 | |
| 1915 | // 12:The output gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 1916 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1917 | // of input and weights scales and zeroPoint equal to 0. |
| 1918 | const ConstTensorPin outputGateBiasPin = |
| 1919 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 12, model, data); |
| 1920 | |
| 1921 | if (!inputToInputWeightsPin.IsValid() || |
| 1922 | !inputToForgetWeightsPin.IsValid() || |
| 1923 | !inputToCellWeightsPin.IsValid() || |
| 1924 | !inputToOutputWeightsPin.IsValid() || |
| 1925 | !recurrentToInputWeightsPin.IsValid() || |
| 1926 | !recurrentToForgetWeightsPin.IsValid() || |
| 1927 | !recurrentToCellWeightsPin.IsValid() || |
| 1928 | !recurrentToOutputWeightsPin.IsValid() || |
| 1929 | !inputGateBiasPin.IsValid() || |
| 1930 | !forgetGateBiasPin.IsValid() || |
| 1931 | !cellBiasPin.IsValid() || |
| 1932 | !outputGateBiasPin.IsValid()) |
| 1933 | { |
| 1934 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 1935 | } |
| 1936 | |
| 1937 | // Outputs: |
| 1938 | // 0: The cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape [numBatches, outputSize] |
| 1939 | // which contains a cell state from the current time step. Tensor is quantized using a quantization range |
| 1940 | // of -2^4, 2^4 * 32767/32768. |
| 1941 | const HalOperand* cellStateOut = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1942 | if (!cellStateOut) |
| 1943 | { |
| 1944 | return Fail("%s: Could not read output 0: cellStateOut", __func__); |
| 1945 | } |
| 1946 | |
| 1947 | // 1: The output: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBathes, outputSize] which |
| 1948 | // contains the output value. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 1949 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 1, model); |
| 1950 | if (!output) |
| 1951 | { |
| 1952 | return Fail("%s: Could not read output 1: output", __func__); |
| 1953 | } |
| 1954 | |
| 1955 | // Inputs |
| 1956 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1957 | const TensorInfo& previousCellStateInInfo = previousCellStateIn.GetTensorInfo(); |
| 1958 | const TensorInfo& previousOutputInInfo = previousOutputIn.GetTensorInfo(); |
| 1959 | |
| 1960 | // Outputs |
| 1961 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 1962 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1963 | |
| 1964 | // Dynamic tensors currently not supported |
| 1965 | if (IsDynamicTensor(cellStateOutInfo) || IsDynamicTensor(outputInfo)) |
| 1966 | { |
| 1967 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1968 | } |
| 1969 | |
| 1970 | QuantizedLstmInputParams params; |
| 1971 | |
| 1972 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 1973 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 1974 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 1975 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 1976 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 1977 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 1978 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 1979 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 1980 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 1981 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 1982 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 1983 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 1984 | |
| 1985 | QuantizedLstmInputParamsInfo paramsInfo; |
| 1986 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 1987 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 1988 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 1989 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 1990 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 1991 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 1992 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 1993 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 1994 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 1995 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 1996 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 1997 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 1998 | |
| 1999 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2000 | armnn::BackendId setBackend; |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2001 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2002 | { |
| 2003 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2004 | IsQuantizedLstmSupported, |
| 2005 | data.m_Backends, |
| 2006 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2007 | setBackend, |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2008 | inputInfo, |
| 2009 | previousCellStateInInfo, |
| 2010 | previousOutputInInfo, |
| 2011 | cellStateOutInfo, |
| 2012 | outputInfo, |
| 2013 | paramsInfo); |
| 2014 | }; |
| 2015 | |
| 2016 | bool isDynamic = false; |
| 2017 | if (!IsDynamicTensor(cellStateOutInfo) && |
| 2018 | !IsDynamicTensor(outputInfo)) |
| 2019 | { |
| 2020 | validateFunc(outputInfo, isSupported); |
| 2021 | } |
| 2022 | else |
| 2023 | { |
| 2024 | isDynamic = true; |
| 2025 | isSupported = AreDynamicTensorsSupported(); |
| 2026 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2027 | |
| 2028 | if (!isSupported) |
| 2029 | { |
| 2030 | return false; |
| 2031 | } |
| 2032 | |
| 2033 | IConnectableLayer* const layer = data.m_Network->AddQuantizedLstmLayer(params, "QuantizedLstm"); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2034 | layer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2035 | input.Connect(layer->GetInputSlot(0)); |
| 2036 | previousCellStateIn.Connect(layer->GetInputSlot(1)); |
| 2037 | previousOutputIn.Connect(layer->GetInputSlot(2)); |
| 2038 | |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2039 | if (!isDynamic) |
| 2040 | { |
| 2041 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 2042 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data)); |
| 2043 | } |
| 2044 | else |
| 2045 | { |
| 2046 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 2047 | SetupAndTrackLayerOutputSlot<HalPolicy>( |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 2048 | operation, 1, *layer, 1, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
Sadik Armagan | baa1f9f | 2020-09-03 10:57:43 +0100 | [diff] [blame] | 2049 | } |
| 2050 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2051 | } |
| 2052 | |
| 2053 | template<typename HalPolicy, |
| 2054 | typename HalOperation = typename HalPolicy::Operation, |
| 2055 | typename HalModel = typename HalPolicy::Model> |
Teresa Charlin | 89cbb3a | 2021-02-11 21:00:47 +0000 | [diff] [blame] | 2056 | bool ConvertReduce(const HalOperation& operation, |
| 2057 | const HalModel& model, |
| 2058 | ConversionData& data, |
| 2059 | ReduceOperation reduceOperation) |
| 2060 | { |
| 2061 | using HalOperand = typename HalPolicy::Operand; |
| 2062 | using HalOperandType = typename HalPolicy::OperandType; |
| 2063 | |
| 2064 | armnn::ReduceDescriptor descriptor; |
| 2065 | descriptor.m_ReduceOperation = reduceOperation; |
| 2066 | |
| 2067 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2068 | if (!input.IsValid()) |
| 2069 | { |
| 2070 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2071 | } |
| 2072 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2073 | |
| 2074 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2075 | if (!output) |
| 2076 | { |
| 2077 | return Fail("%s: Could not read output 0", __func__); |
| 2078 | } |
| 2079 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2080 | |
| 2081 | const HalOperand* axisOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
| 2082 | if (!axisOperand) |
| 2083 | { |
| 2084 | return Fail("%s: Could not read input 1", __func__); |
| 2085 | } |
| 2086 | std::vector<int32_t> axis; |
| 2087 | if (!GetTensorInt32Values<HalPolicy>(*axisOperand, axis, model, data)) |
| 2088 | { |
| 2089 | return Fail("%s: Input 1 has invalid values", __func__); |
| 2090 | } |
| 2091 | |
| 2092 | // Convert the axis to unsigned int and remove duplicates. |
| 2093 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 2094 | std::set<unsigned int> uniqueAxis; |
| 2095 | std::transform(axis.begin(), axis.end(), |
| 2096 | std::inserter(uniqueAxis, uniqueAxis.begin()), |
| 2097 | [rank](int i) -> unsigned int { return (i + rank) % rank; }); |
| 2098 | descriptor.m_vAxis.assign(uniqueAxis.begin(), uniqueAxis.end()); |
| 2099 | |
| 2100 | // Get the "keep dims" flag. |
| 2101 | if (!GetInputScalar<HalPolicy>(operation, 2, HalOperandType::BOOL, descriptor.m_KeepDims, model, data)) |
| 2102 | { |
| 2103 | return Fail("%s: Could not read input 2", __func__); |
| 2104 | } |
| 2105 | |
| 2106 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2107 | armnn::BackendId setBackend; |
Teresa Charlin | 89cbb3a | 2021-02-11 21:00:47 +0000 | [diff] [blame] | 2108 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2109 | { |
| 2110 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2111 | IsReduceSupported, |
| 2112 | data.m_Backends, |
| 2113 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2114 | setBackend, |
Teresa Charlin | 89cbb3a | 2021-02-11 21:00:47 +0000 | [diff] [blame] | 2115 | inputInfo, |
| 2116 | outputInfo, |
| 2117 | descriptor); |
| 2118 | }; |
| 2119 | |
| 2120 | if(!IsDynamicTensor(outputInfo)) |
| 2121 | { |
| 2122 | validateFunc(outputInfo, isSupported); |
| 2123 | } |
| 2124 | else |
| 2125 | { |
| 2126 | isSupported = AreDynamicTensorsSupported(); |
| 2127 | } |
| 2128 | |
| 2129 | if (!isSupported) |
| 2130 | { |
| 2131 | return false; |
| 2132 | } |
| 2133 | |
| 2134 | armnn::IConnectableLayer* const layer = data.m_Network->AddReduceLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2135 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2136 | if (!layer) |
| 2137 | { |
| 2138 | return Fail("%s: Could not add the ReduceLayer", __func__); |
| 2139 | } |
Teresa Charlin | 89cbb3a | 2021-02-11 21:00:47 +0000 | [diff] [blame] | 2140 | input.Connect(layer->GetInputSlot(0)); |
| 2141 | |
| 2142 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2143 | } |
| 2144 | |
| 2145 | template<typename HalPolicy, |
| 2146 | typename HalOperation = typename HalPolicy::Operation, |
| 2147 | typename HalModel = typename HalPolicy::Model> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2148 | bool ConvertResize(const HalOperation& operation, |
| 2149 | const HalModel& model, |
| 2150 | ConversionData& data, |
| 2151 | ResizeMethod resizeMethod) |
| 2152 | { |
| 2153 | using HalOperand = typename HalPolicy::Operand; |
| 2154 | using HalOperandType = typename HalPolicy::OperandType; |
| 2155 | ALOGV("HalPolicy::ConvertResize()"); |
| 2156 | ALOGV("resizeMethod = %s", GetResizeMethodAsCString(resizeMethod)); |
| 2157 | |
| 2158 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2159 | if (!input.IsValid()) |
| 2160 | { |
| 2161 | return Fail("%s: Could not read input 0", __func__); |
| 2162 | } |
| 2163 | |
| 2164 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2165 | if (!output) |
| 2166 | { |
| 2167 | return Fail("%s: Could not read output 0", __func__); |
| 2168 | } |
| 2169 | |
| 2170 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2171 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2172 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2173 | ResizeDescriptor descriptor; |
| 2174 | descriptor.m_Method = resizeMethod; |
| 2175 | descriptor.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 3, model, data); |
| 2176 | |
| 2177 | HalOperandType operandType1; |
| 2178 | HalOperandType operandType2; |
| 2179 | |
| 2180 | if (!GetOperandType<HalPolicy>(operation, 1, model, operandType1) || |
| 2181 | !GetOperandType<HalPolicy>(operation, 2, model, operandType2)) |
| 2182 | { |
| 2183 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2184 | } |
| 2185 | |
| 2186 | if (operandType1 != operandType2) |
| 2187 | { |
| 2188 | return Fail("%s: Operation has invalid inputs. Type of input 1 and 2 should be the same", __func__); |
| 2189 | } |
| 2190 | |
| 2191 | if (operandType1 == HalOperandType::INT32) |
| 2192 | { |
| 2193 | // Case 1: resizing by shape |
| 2194 | int32_t targetWidth = 0; |
| 2195 | int32_t targetHeight = 0; |
| 2196 | |
| 2197 | if (!GetInputInt32<HalPolicy>(operation, 1, targetWidth, model, data) || |
| 2198 | !GetInputInt32<HalPolicy>(operation, 2, targetHeight, model, data)) |
| 2199 | { |
| 2200 | return Fail("%s: Operation has invalid inputs for resizing by shape", __func__); |
| 2201 | } |
| 2202 | |
| 2203 | if (targetWidth < 0 || targetHeight < 0) |
| 2204 | { |
| 2205 | return Fail("%s: Operation has invalid inputs for resizing by shape. " |
| 2206 | "Target width/height cannot be < 0", __func__); |
| 2207 | } |
| 2208 | |
| 2209 | descriptor.m_TargetWidth = static_cast<uint32_t>(targetWidth); |
| 2210 | descriptor.m_TargetHeight = static_cast<uint32_t>(targetHeight); |
| 2211 | } |
| 2212 | else if (operandType1 == HalOperandType::FLOAT32) |
| 2213 | { |
| 2214 | // Case 2: resizing by scale |
| 2215 | float widthScale = 1.0f; |
| 2216 | float heightScale = 1.0f; |
| 2217 | |
| 2218 | if (!GetInputFloat32<HalPolicy>(operation, 1, widthScale, model, data) || |
| 2219 | !GetInputFloat32<HalPolicy>(operation, 2, heightScale, model, data)) |
| 2220 | { |
| 2221 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 2222 | } |
| 2223 | |
| 2224 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 2225 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 2226 | |
| 2227 | float width = inputShape[dataLayoutIndexed.GetWidthIndex()]; |
| 2228 | float height = inputShape[dataLayoutIndexed.GetHeightIndex()]; |
| 2229 | |
| 2230 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 2231 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 2232 | } |
| 2233 | else if (operandType1 == HalOperandType::FLOAT16) |
| 2234 | { |
| 2235 | Half widthScale; |
| 2236 | Half heightScale; |
| 2237 | |
| 2238 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, widthScale, model, data) || |
| 2239 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::FLOAT16, heightScale, model, data)) |
| 2240 | { |
| 2241 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 2242 | } |
| 2243 | |
| 2244 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 2245 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 2246 | |
| 2247 | Half width = static_cast<Half>(inputShape[dataLayoutIndexed.GetWidthIndex()]); |
| 2248 | Half height = static_cast<Half>(inputShape[dataLayoutIndexed.GetHeightIndex()]); |
| 2249 | |
| 2250 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 2251 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 2252 | } |
| 2253 | else |
| 2254 | { |
| 2255 | return Fail("%s: Operand has invalid data type for resizing by scale", __func__); |
| 2256 | } |
| 2257 | |
David Monahan | f057e6f | 2020-06-22 09:55:23 +0100 | [diff] [blame] | 2258 | descriptor.m_AlignCorners = GetOptionalBool<HalPolicy>(operation, 4, model, data); |
| 2259 | descriptor.m_HalfPixelCenters = GetOptionalBool<HalPolicy>(operation, 5, model, data); |
David Monahan | 51e0b13 | 2020-04-20 16:12:06 +0100 | [diff] [blame] | 2260 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2261 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2262 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2263 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2264 | { |
| 2265 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2266 | IsResizeSupported, |
| 2267 | data.m_Backends, |
| 2268 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2269 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2270 | inputInfo, |
| 2271 | outputInfo, |
| 2272 | descriptor); |
| 2273 | }; |
| 2274 | |
| 2275 | if(IsDynamicTensor(outputInfo)) |
| 2276 | { |
| 2277 | isSupported = AreDynamicTensorsSupported(); |
| 2278 | } |
| 2279 | else |
| 2280 | { |
| 2281 | validateFunc(outputInfo, isSupported); |
| 2282 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2283 | |
| 2284 | if (!isSupported) |
| 2285 | { |
| 2286 | return false; |
| 2287 | } |
| 2288 | |
| 2289 | IConnectableLayer* layer = data.m_Network->AddResizeLayer(descriptor); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2290 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2291 | if (!layer) |
| 2292 | { |
| 2293 | return Fail("%s: Could not add the ResizeLayer", __func__); |
| 2294 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2295 | input.Connect(layer->GetInputSlot(0)); |
| 2296 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2297 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2298 | } |
| 2299 | |
| 2300 | template<typename HalPolicy, |
| 2301 | typename HalOperation = typename HalPolicy::Operation, |
| 2302 | typename HalModel = typename HalPolicy::Model> |
| 2303 | bool ConvertSpaceToDepth(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2304 | { |
| 2305 | using HalOperand = typename HalPolicy::Operand; |
| 2306 | using HalOperandType = typename HalPolicy::OperandType; |
| 2307 | |
| 2308 | ALOGV("HalPolicy::ConvertSpaceToDepth()"); |
| 2309 | |
| 2310 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2311 | if (!input.IsValid() ) |
| 2312 | { |
| 2313 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2314 | } |
| 2315 | |
| 2316 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2317 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 2318 | if (rank != 4) |
| 2319 | { |
| 2320 | return Fail("%s: Only inputs with rank 4 are supported", __func__); |
| 2321 | } |
| 2322 | |
| 2323 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2324 | if (!output) |
| 2325 | { |
| 2326 | return Fail("%s: Could not read output 0", __func__); |
| 2327 | } |
| 2328 | |
| 2329 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2330 | |
| 2331 | SpaceToDepthDescriptor desc; |
| 2332 | |
| 2333 | GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, desc.m_BlockSize, model, data); |
| 2334 | |
| 2335 | if (desc.m_BlockSize <= 1) |
| 2336 | { |
Kevin May | 17de62e | 2023-07-31 12:16:04 +0100 | [diff] [blame] | 2337 | return Fail("%s: Block size must be at least 1 in all dimensions", __func__); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2338 | } |
| 2339 | |
| 2340 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 2, model, data); |
| 2341 | |
| 2342 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2343 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2344 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2345 | { |
| 2346 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2347 | IsSpaceToDepthSupported, |
| 2348 | data.m_Backends, |
| 2349 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2350 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2351 | inputInfo, |
| 2352 | outputInfo, |
| 2353 | desc); |
| 2354 | }; |
| 2355 | |
| 2356 | if(IsDynamicTensor(outputInfo)) |
| 2357 | { |
| 2358 | isSupported = AreDynamicTensorsSupported(); |
| 2359 | } |
| 2360 | else |
| 2361 | { |
| 2362 | validateFunc(outputInfo, isSupported); |
| 2363 | } |
| 2364 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2365 | if (!isSupported) |
| 2366 | { |
| 2367 | return false; |
| 2368 | } |
| 2369 | |
| 2370 | IConnectableLayer* const layer = data.m_Network->AddSpaceToDepthLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2371 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2372 | if (!layer) |
| 2373 | { |
Mike Kelly | 1b46d13 | 2021-11-03 11:12:45 +0000 | [diff] [blame] | 2374 | return Fail("%s: Could not add the SpaceToDepthLayer", __func__); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2375 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2376 | input.Connect(layer->GetInputSlot(0)); |
| 2377 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2378 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2379 | } |
| 2380 | |
| 2381 | template<typename HalPolicy, |
| 2382 | typename HalOperation = typename HalPolicy::Operation, |
| 2383 | typename HalModel = typename HalPolicy::Model> |
| 2384 | bool ConvertSoftmax(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2385 | { |
| 2386 | using HalOperand = typename HalPolicy::Operand; |
| 2387 | using HalOperandType = typename HalPolicy::OperandType; |
| 2388 | |
| 2389 | ALOGV("HalPolicy::ConvertSoftmax()"); |
| 2390 | |
| 2391 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2392 | if (!input.IsValid()) |
| 2393 | { |
| 2394 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2395 | } |
| 2396 | |
| 2397 | const HalOperand* outputOperand = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2398 | if (!outputOperand) |
| 2399 | { |
| 2400 | return Fail("%s: Operation has no outputs", __func__); |
| 2401 | } |
| 2402 | |
| 2403 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2404 | |
| 2405 | SoftmaxDescriptor desc; |
Teresa Charlin | bf866e2 | 2020-08-09 23:55:01 +0100 | [diff] [blame] | 2406 | HalOperandType outputType = outputOperand->type; |
| 2407 | |
| 2408 | // Read beta value |
| 2409 | if (outputType == HalOperandType::TENSOR_FLOAT16) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2410 | { |
Teresa Charlin | bf866e2 | 2020-08-09 23:55:01 +0100 | [diff] [blame] | 2411 | Half value; |
| 2412 | |
| 2413 | if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, value, model, data)) |
| 2414 | { |
| 2415 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 2416 | } |
| 2417 | |
| 2418 | desc.m_Beta = static_cast<float>(value); |
| 2419 | } |
| 2420 | else |
| 2421 | { |
| 2422 | if (!GetInputFloat32<HalPolicy>(operation, 1, desc.m_Beta, model, data)) |
| 2423 | { |
| 2424 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 2425 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2426 | } |
| 2427 | |
| 2428 | if (operation.inputs.size() > 2 && !GetInputScalar<HalPolicy>(operation, |
Teresa Charlin | bf866e2 | 2020-08-09 23:55:01 +0100 | [diff] [blame] | 2429 | 2, |
| 2430 | HalOperandType::INT32, |
| 2431 | desc.m_Axis, |
| 2432 | model, |
| 2433 | data)) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2434 | { |
| 2435 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2436 | } |
| 2437 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2438 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2439 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2440 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2441 | { |
| 2442 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2443 | IsSoftmaxSupported, |
| 2444 | data.m_Backends, |
| 2445 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2446 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2447 | input.GetTensorInfo(), |
| 2448 | outputInfo, |
| 2449 | desc); |
| 2450 | }; |
| 2451 | |
| 2452 | if(IsDynamicTensor(outputInfo)) |
| 2453 | { |
| 2454 | isSupported = AreDynamicTensorsSupported(); |
| 2455 | } |
| 2456 | else |
| 2457 | { |
| 2458 | validateFunc(outputInfo, isSupported); |
| 2459 | } |
| 2460 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2461 | if (!isSupported) |
| 2462 | { |
| 2463 | return false; |
| 2464 | } |
| 2465 | |
| 2466 | IConnectableLayer* layer = data.m_Network->AddSoftmaxLayer(desc); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2467 | layer->SetBackendId(setBackend); |
Mike Kelly | e2d611e | 2021-10-14 12:35:58 +0100 | [diff] [blame] | 2468 | if (!layer) |
| 2469 | { |
| 2470 | return Fail("%s: Could not add the SoftmaxLayer", __func__); |
| 2471 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2472 | input.Connect(layer->GetInputSlot(0)); |
| 2473 | |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 2474 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data, nullptr, validateFunc); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2475 | } |
| 2476 | |
| 2477 | template<typename HalPolicy, |
Mike Kelly | d261ffe | 2023-08-04 13:17:23 +0100 | [diff] [blame^] | 2478 | typename Operation = typename HalPolicy::Operation, |
| 2479 | typename Model = typename HalPolicy::Model> |
| 2480 | bool ConvertSplit(const Operation& operation, const Model& model, ConversionData& data) |
| 2481 | { |
| 2482 | using HalOperand = typename HalPolicy::Operand; |
| 2483 | using HalOperandType = typename HalPolicy::OperandType; |
| 2484 | |
| 2485 | if (operation.inputs.size() != 3) |
| 2486 | { |
| 2487 | return Fail("%s: Optional inputs are not supported expected 3 was %i", __func__, operation.inputs.size()); |
| 2488 | } |
| 2489 | |
| 2490 | // 0: An n-D tensor to split. |
| 2491 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2492 | // 1: An ANEURALNETWORKS_INT32 scalar specifying the axis along which to split. |
| 2493 | int32_t axis = 0; |
| 2494 | // 2: An ANEURALNETWORKS_INT32 scalar indicating the number of splits along given axis. |
| 2495 | // Must evenly divide axis size. |
| 2496 | int32_t numSplits = 0; |
| 2497 | |
| 2498 | if (!input.IsValid() || |
| 2499 | !GetInputScalar<HalPolicy>(operation, 1, HalOperandType::INT32, axis, model, data) || |
| 2500 | !GetInputScalar<HalPolicy>(operation, 2, HalOperandType::INT32, numSplits, model, data)) |
| 2501 | { |
| 2502 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2503 | } |
| 2504 | |
| 2505 | // If number of splits is <= zero, return false. |
| 2506 | if (numSplits <= 0) |
| 2507 | { |
| 2508 | return Fail("%s: Number of splits must be greater than zero", __func__); |
| 2509 | } |
| 2510 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2511 | unsigned int inputDimSize = inputInfo.GetNumDimensions(); |
| 2512 | int32_t inputDimensions = static_cast<int32_t>(inputDimSize); |
| 2513 | |
| 2514 | if (axis < -inputDimensions || axis >= inputDimensions) |
| 2515 | { |
| 2516 | // The axis for a tensor with n dimensions must be between -n and n-1 |
| 2517 | // E.g. Rank 4 tensor can have axis in range [-4, 3) |
| 2518 | // -1 == 3, -2 == 2, -3 == 1, -4 == 0 |
| 2519 | return Fail("%s: Operation has invalid axis %i. Axis must be in range [-n, n-1]", __func__, axis); |
| 2520 | } |
| 2521 | auto splitDim = armnnUtils::GetUnsignedAxis(inputDimSize, axis); |
| 2522 | |
| 2523 | if (inputDimSize > MaxNumOfTensorDimensions) |
| 2524 | { |
| 2525 | return Fail("%s: The number of dimensions %i for split operation cannot be greater than %i", |
| 2526 | __func__, inputInfo.GetNumDimensions(), MaxNumOfTensorDimensions); |
| 2527 | } |
| 2528 | std::vector<uint32_t> splitterDimSizes(inputDimSize); |
| 2529 | |
| 2530 | // Add current input shape to splitterDimSizes |
| 2531 | for (uint32_t i = 0; i < inputDimSize; ++i) |
| 2532 | { |
| 2533 | splitterDimSizes[i] = inputInfo.GetShape()[i]; |
| 2534 | } |
| 2535 | |
| 2536 | if (splitterDimSizes[splitDim] % numSplits != 0) |
| 2537 | { |
| 2538 | return Fail("%s: The number of splits %i must evenly divide the dimension %i", |
| 2539 | __func__, numSplits, splitterDimSizes[splitDim]); |
| 2540 | } |
| 2541 | splitterDimSizes[splitDim] /= numSplits; |
| 2542 | |
| 2543 | ViewsDescriptor descriptor(numSplits, inputDimSize); |
| 2544 | |
| 2545 | for (int32_t i = 0; i < numSplits; ++i) |
| 2546 | { |
| 2547 | // Set the size of the views. |
| 2548 | for (uint32_t dimIdx = 0; dimIdx < splitterDimSizes.size(); ++dimIdx) |
| 2549 | { |
| 2550 | descriptor.SetViewSize(i, dimIdx, splitterDimSizes[dimIdx]); |
| 2551 | } |
| 2552 | descriptor.SetViewOriginCoord(i, splitDim, splitterDimSizes[splitDim] * i); |
| 2553 | } |
| 2554 | |
| 2555 | std::vector<TensorInfo> outputInfos; |
| 2556 | for (int32_t i = 0; i < numSplits; ++i) |
| 2557 | { |
| 2558 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, i, model); |
| 2559 | if (!output) |
| 2560 | { |
| 2561 | return Fail("%s: Could not read output %i", __func__, i); |
| 2562 | } |
| 2563 | |
| 2564 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2565 | outputInfos.template emplace_back(outputInfo); |
| 2566 | } |
| 2567 | std::vector<std::reference_wrapper<TensorInfo>> splitterOutputInfos(outputInfos.begin(), outputInfos.end()); |
| 2568 | bool isSupported = false; |
| 2569 | armnn::BackendId setBackend; |
| 2570 | |
| 2571 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2572 | IsSplitterSupported, |
| 2573 | data.m_Backends, |
| 2574 | isSupported, |
| 2575 | setBackend, |
| 2576 | inputInfo, |
| 2577 | splitterOutputInfos, |
| 2578 | descriptor); |
| 2579 | |
| 2580 | if (!isSupported) |
| 2581 | { |
| 2582 | return Fail("%s: Layer is not supported", __func__); |
| 2583 | } |
| 2584 | |
| 2585 | for (int32_t i = 0; i < numSplits; ++i) |
| 2586 | { |
| 2587 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, i, model); |
| 2588 | if (!output) |
| 2589 | { |
| 2590 | return Fail("%s: Could not read output %i", __func__, i); |
| 2591 | } |
| 2592 | |
| 2593 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2594 | if (IsDynamicTensor(outputInfo)) |
| 2595 | { |
| 2596 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 2597 | } |
| 2598 | } |
| 2599 | |
| 2600 | IConnectableLayer* layer = data.m_Network->AddSplitterLayer(descriptor, "Split"); |
| 2601 | if (!layer) |
| 2602 | { |
| 2603 | return Fail("%s: could not add the Layer", __func__); |
| 2604 | } |
| 2605 | input.Connect(layer->GetInputSlot(0)); |
| 2606 | |
| 2607 | auto validateFunc = [&](const armnn::TensorInfo&, bool& isSupported) |
| 2608 | { |
| 2609 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2610 | IsSplitterSupported, |
| 2611 | data.m_Backends, |
| 2612 | isSupported, |
| 2613 | setBackend, |
| 2614 | inputInfo, |
| 2615 | splitterOutputInfos, |
| 2616 | descriptor); |
| 2617 | }; |
| 2618 | |
| 2619 | for (int32_t i = 0; i < numSplits; ++i) |
| 2620 | { |
| 2621 | bool ok = SetupAndTrackLayerOutputSlot<HalPolicy>(operation, i, *layer, model, data, nullptr, validateFunc); |
| 2622 | |
| 2623 | if (!ok) |
| 2624 | { |
| 2625 | return false; |
| 2626 | } |
| 2627 | } |
| 2628 | return true; |
| 2629 | } |
| 2630 | |
| 2631 | template<typename HalPolicy, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2632 | typename HalOperation = typename HalPolicy::Operation, |
| 2633 | typename HalModel = typename HalPolicy::Model> |
| 2634 | bool ConvertLstm(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 2635 | { |
| 2636 | using HalOperand = typename HalPolicy::Operand; |
| 2637 | using HalOperandType = typename HalPolicy::OperandType; |
| 2638 | |
| 2639 | ALOGV("HalPolicy::ConvertLstm()"); |
| 2640 | |
| 2641 | // Inputs: |
| 2642 | // 00: The input: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where |
| 2643 | // “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input. |
| 2644 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 2645 | if (!input.IsValid()) |
| 2646 | { |
| 2647 | return Fail("%s: Could not read input 0: input", __func__); |
| 2648 | } |
| 2649 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2650 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 18, model, data); |
| 2651 | if (!outputStateIn.IsValid()) |
| 2652 | { |
| 2653 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 2654 | } |
| 2655 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2656 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 19, model, data); |
| 2657 | if (!cellStateIn.IsValid()) |
| 2658 | { |
| 2659 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 2660 | } |
| 2661 | |
| 2662 | // Get the mandatory input tensors: |
| 2663 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2664 | // [num_units, input_size]. |
| 2665 | const ConstTensorPin inputToForgetWeightsPin = |
| 2666 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 2)); |
| 2667 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2668 | // [num_units, input_size]. |
| 2669 | const ConstTensorPin inputToCellWeightsPin = |
| 2670 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 3)); |
| 2671 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2672 | // [num_units, input_size]. |
| 2673 | const ConstTensorPin inputToOutputWeightsPin = |
| 2674 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 4)); |
| 2675 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2676 | // [num_units, output_size]. |
| 2677 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 2678 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 6)); |
| 2679 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2680 | // [num_units, output_size]. |
| 2681 | const ConstTensorPin recurrentToCellWeightsPin = |
| 2682 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 7)); |
| 2683 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2684 | // [num_units, output_size]. |
| 2685 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 2686 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 8)); |
| 2687 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2688 | const ConstTensorPin forgetGateBiasPin = |
| 2689 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 13, model, data); |
| 2690 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2691 | const ConstTensorPin cellBiasPin = |
| 2692 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 14, model, data); |
| 2693 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2694 | const ConstTensorPin outputGateBiasPin = |
| 2695 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 15, model, data); |
| 2696 | |
| 2697 | if (!inputToForgetWeightsPin.IsValid() || |
| 2698 | !inputToCellWeightsPin.IsValid() || |
| 2699 | !inputToOutputWeightsPin.IsValid() || |
| 2700 | !recurrentToForgetWeightsPin.IsValid() || |
| 2701 | !recurrentToCellWeightsPin.IsValid() || |
| 2702 | !recurrentToOutputWeightsPin.IsValid() || |
| 2703 | !forgetGateBiasPin.IsValid() || |
| 2704 | !cellBiasPin.IsValid() || |
| 2705 | !outputGateBiasPin.IsValid()) |
| 2706 | { |
| 2707 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2708 | } |
| 2709 | |
| 2710 | // Get the optional input tensors: |
| 2711 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2712 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 2713 | const ConstTensorPin inputToInputWeightsPin = |
| 2714 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 1, true)); |
| 2715 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2716 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 2717 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 2718 | const ConstTensorPin recurrentToInputWeightsPin = |
| 2719 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 5, true)); |
| 2720 | // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2721 | const ConstTensorPin cellToInputWeightsPin = |
| 2722 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 9, true)); |
| 2723 | // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2724 | const ConstTensorPin cellToForgetWeightsPin = |
| 2725 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 10, true)); |
| 2726 | // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2727 | const ConstTensorPin cellToOutputWeightsPin = |
| 2728 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 11, true)); |
| 2729 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2730 | const ConstTensorPin inputGateBiasPin = |
| 2731 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2732 | 12, |
| 2733 | model, |
| 2734 | data, |
| 2735 | g_DontPermute, |
| 2736 | nullptr, |
| 2737 | true); |
| 2738 | |
| 2739 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2740 | // [output_size, num_units]. |
| 2741 | const ConstTensorPin projectionWeightsPin = |
| 2742 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 16, true)); |
| 2743 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size]. |
| 2744 | const ConstTensorPin projectionBiasPin = |
| 2745 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2746 | 17, |
| 2747 | model, |
| 2748 | data, |
| 2749 | g_DontPermute, |
| 2750 | nullptr, |
| 2751 | true); |
| 2752 | |
| 2753 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 2754 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 2755 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 2756 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 2757 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 2758 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 2759 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 2760 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 2761 | { |
| 2762 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2763 | } |
| 2764 | |
| 2765 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 2766 | // 20: The activation function: A value indicating the activation function: |
| 2767 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 2768 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 2769 | // If set to 0.0 then clipping is disabled. |
| 2770 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 2771 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
Cathal Corbett | 3fe3a54 | 2022-05-25 09:35:37 +0100 | [diff] [blame] | 2772 | ActivationFn activation = ActivationFn::kActivationNone; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2773 | float cellClip; |
| 2774 | float projClip; |
| 2775 | if (!GetInputActivationFunctionFromTensor<HalPolicy>(operation, 20, activation, model, data) || |
| 2776 | !GetInputScalar<HalPolicy>(operation, 21, HalOperandType::FLOAT32, cellClip, model, data) || |
| 2777 | !GetInputScalar<HalPolicy>(operation, 22, HalOperandType::FLOAT32, projClip, model, data)) |
| 2778 | { |
| 2779 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 2780 | } |
| 2781 | |
| 2782 | // Get the normalization tensors |
| 2783 | // 23: The input layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2784 | // Used to rescale normalized inputs to activation at input gate. |
| 2785 | const ConstTensorPin inputLayerNormWeightsPin |
| 2786 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 23, true)); |
| 2787 | |
| 2788 | // 24: The forget layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2789 | // Used to rescale normalized inputs to activation at forget gate. |
| 2790 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 2791 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2792 | 24, |
| 2793 | model, |
| 2794 | data, |
| 2795 | g_DontPermute, |
| 2796 | nullptr, |
| 2797 | true); |
| 2798 | |
| 2799 | // 25: The cell layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2800 | // Used to rescale normalized inputs to activation at cell gate. |
| 2801 | const ConstTensorPin cellLayerNormWeightsPin = |
| 2802 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2803 | 25, |
| 2804 | model, |
| 2805 | data, |
| 2806 | g_DontPermute, |
| 2807 | nullptr, |
| 2808 | true); |
| 2809 | |
| 2810 | // 26: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2811 | // Used to rescale normalized inputs to activation at output gate. |
| 2812 | const ConstTensorPin outputLayerNormWeightsPin = |
| 2813 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 2814 | 26, |
| 2815 | model, |
| 2816 | data, |
| 2817 | g_DontPermute, |
| 2818 | nullptr, |
| 2819 | true); |
| 2820 | |
| 2821 | // Outputs: |
| 2822 | // 00: The scratch buffer: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] |
| 2823 | // with CIFG, or [batch_size, num_units * 3] without CIFG. |
| 2824 | const HalOperand* scratchBuffer = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 2825 | if (!scratchBuffer) |
| 2826 | { |
| 2827 | return Fail("%s: Could not read output 0: scratchBuffer", __func__); |
| 2828 | } |
| 2829 | // 01: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2830 | const HalOperand* outputStateOut = GetOutputOperand<HalPolicy>(operation, 1, model); |
| 2831 | if (!outputStateOut) |
| 2832 | { |
| 2833 | return Fail("%s: Could not read output 1: outputStateOut", __func__); |
| 2834 | } |
| 2835 | // 02: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2836 | const HalOperand* cellStateOut = GetOutputOperand<HalPolicy>(operation, 2, model); |
| 2837 | if (!cellStateOut) |
| 2838 | { |
| 2839 | return Fail("%s: Could not read output 2: cellStateOut", __func__); |
| 2840 | } |
| 2841 | // 03: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is |
| 2842 | // effectively the same as the current “output state (out)” value. |
| 2843 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 3, model); |
| 2844 | if (!output) |
| 2845 | { |
| 2846 | return Fail("%s: Could not read output 3: output", __func__); |
| 2847 | } |
| 2848 | |
| 2849 | // set the params structure for the AddLstmLayer call |
| 2850 | LstmInputParams params; |
| 2851 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 2852 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 2853 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 2854 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 2855 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 2856 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 2857 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 2858 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 2859 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 2860 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 2861 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 2862 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 2863 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 2864 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 2865 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 2866 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 2867 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 2868 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 2869 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 2870 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 2871 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 2872 | |
| 2873 | // set the layer descriptor |
| 2874 | LstmDescriptor desc; |
| 2875 | desc.m_ActivationFunc = activation; |
| 2876 | desc.m_ClippingThresCell = cellClip; |
| 2877 | desc.m_ClippingThresProj = projClip; |
| 2878 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 2879 | params.m_RecurrentToInputWeights == nullptr || |
| 2880 | params.m_InputGateBias == nullptr); |
| 2881 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 2882 | params.m_CellToOutputWeights != nullptr); |
| 2883 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 2884 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 2885 | params.m_ForgetLayerNormWeights != nullptr || |
| 2886 | params.m_CellLayerNormWeights != nullptr || |
| 2887 | params.m_OutputLayerNormWeights != nullptr); |
| 2888 | |
| 2889 | // validate the optional input groups |
| 2890 | if (desc.m_CifgEnabled && |
| 2891 | (params.m_InputToInputWeights != nullptr || |
| 2892 | params.m_RecurrentToInputWeights != nullptr || |
| 2893 | params.m_InputGateBias != nullptr)) |
| 2894 | { |
| 2895 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 2896 | " and input gate bias must be provided", __func__); |
| 2897 | } |
| 2898 | |
| 2899 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 2900 | { |
| 2901 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 2902 | } |
| 2903 | |
| 2904 | if (desc.m_PeepholeEnabled && |
| 2905 | (params.m_CellToForgetWeights == nullptr || |
| 2906 | params.m_CellToOutputWeights == nullptr || |
| 2907 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 2908 | { |
| 2909 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 2910 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 2911 | } |
| 2912 | |
| 2913 | if (desc.m_LayerNormEnabled && |
| 2914 | (params.m_ForgetLayerNormWeights == nullptr || |
| 2915 | params.m_CellLayerNormWeights == nullptr || |
| 2916 | params.m_OutputLayerNormWeights == nullptr || |
| 2917 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 2918 | { |
| 2919 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 2920 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 2921 | } |
| 2922 | |
| 2923 | // Check if the layer is supported |
| 2924 | // Inputs |
| 2925 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2926 | const TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 2927 | const TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
| 2928 | |
| 2929 | // Outputs |
| 2930 | const TensorInfo& scratchBufferInfo = GetTensorInfoForOperand(*scratchBuffer); |
| 2931 | const TensorInfo& outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); |
| 2932 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 2933 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2934 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 2935 | // Basic parameters |
| 2936 | LstmInputParamsInfo paramsInfo; |
| 2937 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 2938 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 2939 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 2940 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 2941 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 2942 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 2943 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 2944 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 2945 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 2946 | |
| 2947 | // Optional parameters |
| 2948 | if (!desc.m_CifgEnabled) |
| 2949 | { |
| 2950 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 2951 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 2952 | if (params.m_CellToInputWeights != nullptr) |
| 2953 | { |
| 2954 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 2955 | } |
| 2956 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 2957 | } |
| 2958 | |
| 2959 | if (desc.m_ProjectionEnabled) |
| 2960 | { |
| 2961 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 2962 | if (params.m_ProjectionBias != nullptr) |
| 2963 | { |
| 2964 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 2965 | } |
| 2966 | } |
| 2967 | |
| 2968 | if (desc.m_PeepholeEnabled) |
| 2969 | { |
| 2970 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 2971 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 2972 | } |
| 2973 | |
| 2974 | if (desc.m_LayerNormEnabled) |
| 2975 | { |
| 2976 | if(!desc.m_CifgEnabled) |
| 2977 | { |
| 2978 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 2979 | } |
| 2980 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 2981 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 2982 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 2983 | } |
| 2984 | |
| 2985 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2986 | armnn::BackendId setBackend; |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 2987 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2988 | { |
| 2989 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2990 | IsLstmSupported, |
| 2991 | data.m_Backends, |
| 2992 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 2993 | setBackend, |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 2994 | inputInfo, |
| 2995 | outputStateInInfo, |
| 2996 | cellStateInInfo, |
| 2997 | scratchBufferInfo, |
| 2998 | outputStateOutInfo, |
| 2999 | cellStateOutInfo, |
| 3000 | outputInfo, |
| 3001 | desc, |
| 3002 | paramsInfo); |
| 3003 | }; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 3004 | |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 3005 | bool isDynamic = false; |
| 3006 | if (!IsDynamicTensor(outputStateOutInfo) && |
| 3007 | !IsDynamicTensor(scratchBufferInfo) && |
| 3008 | !IsDynamicTensor(cellStateOutInfo) && |
| 3009 | !IsDynamicTensor(outputInfo)) |
| 3010 | { |
| 3011 | validateFunc(outputInfo, isSupported); |
| 3012 | } |
| 3013 | else |
| 3014 | { |
| 3015 | isDynamic = true; |
| 3016 | isSupported = AreDynamicTensorsSupported(); |
| 3017 | } |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 3018 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3019 | if (!isSupported) |
| 3020 | { |
| 3021 | return false; |
| 3022 | } |
| 3023 | |
| 3024 | // Add the layer |
| 3025 | IConnectableLayer* layer = data.m_Network->AddLstmLayer(desc, params, "Lstm"); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3026 | layer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3027 | |
| 3028 | input.Connect(layer->GetInputSlot(0)); |
| 3029 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 3030 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 3031 | |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 3032 | if (!isDynamic) |
| 3033 | { |
| 3034 | return ( |
| 3035 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 3036 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data) && |
| 3037 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 2, *layer, 2, model, data) && |
| 3038 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 3, *layer, 3, model, data)); |
| 3039 | } |
| 3040 | else |
| 3041 | { |
| 3042 | return ( |
| 3043 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 3044 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data) && |
| 3045 | SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 2, *layer, 2, model, data) && |
| 3046 | SetupAndTrackLayerOutputSlot<HalPolicy>( |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 3047 | operation, 3, *layer, 3, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
Sadik Armagan | dbda4b7 | 2020-09-03 11:33:07 +0100 | [diff] [blame] | 3048 | } |
| 3049 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3050 | } |
| 3051 | |
| 3052 | template<typename HalPolicy, |
| 3053 | typename HalOperation = typename HalPolicy::Operation, |
| 3054 | typename HalModel = typename HalPolicy::Model> |
Teresa Charlin | 7f0ff16 | 2023-07-24 23:42:10 +0100 | [diff] [blame] | 3055 | bool ConvertTile(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 3056 | { |
| 3057 | using HalOperand = typename HalPolicy::Operand; |
| 3058 | |
| 3059 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 3060 | |
| 3061 | if (!input.IsValid()) |
| 3062 | { |
| 3063 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3064 | } |
| 3065 | |
| 3066 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 3067 | |
| 3068 | if (!output) |
| 3069 | { |
| 3070 | return Fail("%s: Could not read output", __func__); |
| 3071 | } |
| 3072 | |
| 3073 | const HalOperand* multiplesOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
| 3074 | if (!multiplesOperand) |
| 3075 | { |
| 3076 | return Fail("%s: Could not read input 1", __func__); |
| 3077 | } |
| 3078 | std::vector<int32_t> multiples; |
| 3079 | if (!GetTensorInt32Values<HalPolicy>(*multiplesOperand, multiples, model, data)) |
| 3080 | { |
| 3081 | return Fail("%s: Input 1 has invalid values", __func__); |
| 3082 | } |
| 3083 | // Convert the multiples from int to unsigned int, |
| 3084 | // as values are always going to be positive despite the data type being integer. |
| 3085 | TileDescriptor descriptor; |
| 3086 | descriptor.m_Multiples.assign(multiples.begin(), multiples.end()); |
| 3087 | |
| 3088 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3089 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3090 | |
| 3091 | bool isSupported = false; |
| 3092 | armnn::BackendId setBackend; |
| 3093 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3094 | { |
| 3095 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3096 | IsTileSupported, |
| 3097 | data.m_Backends, |
| 3098 | isSupported, |
| 3099 | setBackend, |
| 3100 | inputInfo, |
| 3101 | outputInfo, |
| 3102 | descriptor); |
| 3103 | }; |
| 3104 | |
| 3105 | if(IsDynamicTensor(outputInfo)) |
| 3106 | { |
| 3107 | isSupported = AreDynamicTensorsSupported(); |
| 3108 | } |
| 3109 | else |
| 3110 | { |
| 3111 | validateFunc(outputInfo, isSupported); |
| 3112 | } |
| 3113 | if (!isSupported) |
| 3114 | { |
| 3115 | return false; |
| 3116 | } |
| 3117 | |
| 3118 | IConnectableLayer* tileLayer = data.m_Network->AddTileLayer(descriptor); |
| 3119 | if (!tileLayer) |
| 3120 | { |
| 3121 | return Fail("%s: AddTileLayer failed", __func__); |
| 3122 | } |
| 3123 | tileLayer->SetBackendId(setBackend); |
| 3124 | |
| 3125 | input.Connect(tileLayer->GetInputSlot(0)); |
| 3126 | |
| 3127 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *tileLayer, model, data, nullptr, validateFunc); |
| 3128 | } |
| 3129 | |
| 3130 | template<typename HalPolicy, |
| 3131 | typename HalOperation = typename HalPolicy::Operation, |
| 3132 | typename HalModel = typename HalPolicy::Model> |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3133 | bool ConvertTransposeConv2d(const HalOperation& operation, const HalModel& model, ConversionData& data) |
| 3134 | { |
| 3135 | using HalOperand = typename HalPolicy::Operand; |
| 3136 | using HalOperandType = typename HalPolicy::OperandType; |
| 3137 | |
| 3138 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 3139 | |
| 3140 | if (!input.IsValid()) |
| 3141 | { |
| 3142 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3143 | } |
| 3144 | |
| 3145 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 3146 | |
| 3147 | if (!output) |
| 3148 | { |
| 3149 | return Fail("%s: Could not read output 0", __func__); |
| 3150 | } |
| 3151 | |
| 3152 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3153 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3154 | |
| 3155 | // ArmNN does not currently support non-fixed weights or bias |
| 3156 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 3157 | const HalOperand* weightsOperand = GetInputOperand<HalPolicy>(operation, 1, model); |
| 3158 | |
| 3159 | if (weightsOperand == nullptr) |
| 3160 | { |
| 3161 | return Fail("%s: Operand is invalid", __func__); |
| 3162 | } |
| 3163 | TransposeConvolution2dDescriptor desc; |
| 3164 | desc.m_DataLayout = DataLayout::NHWC; |
| 3165 | |
| 3166 | // Determine whether padding is implicit or explicit |
| 3167 | bool implicitPadding = operation.inputs.size() == 9; |
| 3168 | |
| 3169 | if (implicitPadding ) |
| 3170 | { |
| 3171 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 8, model, data); |
| 3172 | } |
| 3173 | else |
| 3174 | { |
| 3175 | desc.m_DataLayout = OptionalDataLayout<HalPolicy>(operation, 10, model, data); |
| 3176 | } |
| 3177 | |
| 3178 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 3179 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 3180 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 3181 | |
| 3182 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
| 3183 | |
| 3184 | // The shape of the weight is [depth_out, filter_height, filter_width, depth_in]. |
| 3185 | // We have to permute it to OIHW if the data layout is NCHW. |
| 3186 | const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ? |
| 3187 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, |
| 3188 | model, data, OHWIToOIHW) : |
| 3189 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data); |
| 3190 | |
| 3191 | // Bias is a 1D tensor |
| 3192 | const ConstTensorPin biasPin = |
| 3193 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); |
| 3194 | |
| 3195 | if (!weightsPin.IsValid()) |
| 3196 | { |
| 3197 | return Fail("%s: Operation has invalid weights", __func__); |
| 3198 | } |
| 3199 | |
| 3200 | if (!biasPin.IsValid()) |
| 3201 | { |
| 3202 | return Fail("%s: Operation has invalid biases", __func__); |
| 3203 | } |
| 3204 | |
| 3205 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 3206 | ConstTensor bias = biasPin.GetConstTensor(); |
| 3207 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 3208 | |
| 3209 | ActivationFn activation; |
| 3210 | |
| 3211 | if (implicitPadding) |
| 3212 | { |
| 3213 | int32_t strideX{0}; |
| 3214 | int32_t strideY{0}; |
| 3215 | int32_t padLeft{0}; |
| 3216 | int32_t padRight{0}; |
| 3217 | int32_t padTop{0}; |
| 3218 | int32_t padBottom{0}; |
| 3219 | |
| 3220 | android::nn::PaddingScheme paddingScheme; |
| 3221 | if (!GetInputPaddingScheme<HalPolicy>(operation, 4, paddingScheme, model, data) || |
| 3222 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, strideX, model, data) || |
| 3223 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, strideY, model, data) || |
| 3224 | !GetInputActivationFunction<HalPolicy>(operation, 7, activation, model, data)) |
| 3225 | { |
| 3226 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 3227 | } |
| 3228 | |
| 3229 | const uint32_t kernelX = weights.GetShape()[widthIndex]; |
| 3230 | const uint32_t kernelY = weights.GetShape()[heightIndex]; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3231 | |
Colm Donelan | 8f3d33e | 2020-07-06 15:41:43 +0100 | [diff] [blame] | 3232 | // If output shape has been specified as a parameter then extract it and make it available. |
| 3233 | const HalOperand* outputShapeOperand = GetInputOperand<HalPolicy>(operation, 3, model, false); |
| 3234 | std::vector<int32_t> outputShape; |
| 3235 | if ((outputShapeOperand) && (GetTensorInt32Values<HalPolicy>(*outputShapeOperand, outputShape, model, data))) |
| 3236 | { |
| 3237 | // Change from signed to unsigned int to store in TransposeConvolution2dDescriptor. |
| 3238 | for (int dimension : outputShape) |
| 3239 | { |
| 3240 | desc.m_OutputShape.push_back(static_cast<unsigned int>(dimension)); |
| 3241 | } |
| 3242 | desc.m_OutputShapeEnabled = true; |
| 3243 | } |
| 3244 | |
Finn Williams | 8fe50c6 | 2020-10-09 15:52:57 +0100 | [diff] [blame] | 3245 | uint32_t outputX; |
| 3246 | uint32_t outputY; |
| 3247 | |
| 3248 | if (IsDynamicTensor(outputInfo)) |
| 3249 | { |
| 3250 | if (outputShape.size() == 0) |
| 3251 | { |
| 3252 | return Fail("%s: Padding sizes cannot be inferred", __func__); |
| 3253 | } |
| 3254 | |
| 3255 | outputX = outputShape[widthIndex]; |
| 3256 | outputY = outputShape[heightIndex]; |
| 3257 | } |
| 3258 | else |
| 3259 | { |
| 3260 | outputX = outputInfo.GetShape()[widthIndex]; |
| 3261 | outputY = outputInfo.GetShape()[heightIndex]; |
| 3262 | } |
| 3263 | |
| 3264 | CalcPaddingTransposeConv(outputX, kernelX, strideX, padLeft, padRight, paddingScheme); |
| 3265 | CalcPaddingTransposeConv(outputY, kernelY, strideY, padTop, padBottom, paddingScheme); |
| 3266 | |
| 3267 | // NOTE: The Android NN API allows for negative padding values in TransposeConv2d, |
| 3268 | // but Arm NN only supports values >= 0 |
| 3269 | if (padLeft < 0 || padRight < 0 || padTop < 0 || padBottom < 0) |
| 3270 | { |
| 3271 | return Fail("%s: Negative padding values are not supported", __func__); |
| 3272 | } |
| 3273 | |
Matthew Sloyan | 9b088d9 | 2020-09-14 15:12:55 +0100 | [diff] [blame] | 3274 | desc.m_StrideX = armnn::numeric_cast<uint32_t>(strideX); |
| 3275 | desc.m_StrideY = armnn::numeric_cast<uint32_t>(strideY); |
| 3276 | desc.m_PadLeft = armnn::numeric_cast<uint32_t>(padLeft); |
| 3277 | desc.m_PadRight = armnn::numeric_cast<uint32_t>(padRight); |
| 3278 | desc.m_PadTop = armnn::numeric_cast<uint32_t>(padTop); |
| 3279 | desc.m_PadBottom = armnn::numeric_cast<uint32_t>(padBottom); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3280 | } |
| 3281 | else if (operation.inputs.size() == 11) |
| 3282 | { |
| 3283 | // explicit padding |
| 3284 | if (!GetInputScalar<HalPolicy>(operation, 3, HalOperandType::INT32, desc.m_PadLeft, model, data) || |
| 3285 | !GetInputScalar<HalPolicy>(operation, 4, HalOperandType::INT32, desc.m_PadRight, model, data) || |
| 3286 | !GetInputScalar<HalPolicy>(operation, 5, HalOperandType::INT32, desc.m_PadTop, model, data) || |
| 3287 | !GetInputScalar<HalPolicy>(operation, 6, HalOperandType::INT32, desc.m_PadBottom, model, data) || |
| 3288 | !GetInputScalar<HalPolicy>(operation, 7, HalOperandType::INT32, desc.m_StrideX, model, data) || |
| 3289 | !GetInputScalar<HalPolicy>(operation, 8, HalOperandType::INT32, desc.m_StrideY, model, data) || |
| 3290 | !GetInputActivationFunction<HalPolicy>(operation, 9, activation, model, data)) |
| 3291 | { |
| 3292 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 3293 | } |
| 3294 | } |
| 3295 | else |
| 3296 | { |
| 3297 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 3298 | } |
| 3299 | |
| 3300 | desc.m_BiasEnabled = true; |
| 3301 | Optional<TensorInfo> biases(bias.GetInfo()); |
| 3302 | |
| 3303 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3304 | armnn::BackendId setBackend; |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 3305 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3306 | { |
| 3307 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3308 | IsTransposeConvolution2dSupported, |
| 3309 | data.m_Backends, |
| 3310 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3311 | setBackend, |
Teresa Charlin | 4bd9a74 | 2020-08-12 12:58:50 +0100 | [diff] [blame] | 3312 | inputInfo, |
| 3313 | outputInfo, |
| 3314 | desc, |
| 3315 | weights.GetInfo(), |
| 3316 | biases); |
| 3317 | }; |
| 3318 | |
| 3319 | if(IsDynamicTensor(outputInfo)) |
| 3320 | { |
| 3321 | isSupported = AreDynamicTensorsSupported(); |
| 3322 | } |
| 3323 | else |
| 3324 | { |
| 3325 | validateFunc(outputInfo, isSupported); |
| 3326 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3327 | if (!isSupported) |
| 3328 | { |
| 3329 | return false; |
| 3330 | } |
| 3331 | |
| 3332 | IConnectableLayer* startLayer = |
| 3333 | data.m_Network->AddTransposeConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3334 | startLayer->SetBackendId(setBackend); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3335 | if (!startLayer) |
| 3336 | { |
| 3337 | return Fail("%s: AddTransposeConvolution2dLayer failed", __func__); |
| 3338 | } |
| 3339 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3340 | input.Connect(startLayer->GetInputSlot(0)); |
| 3341 | |
Kevin May | fcf2a15 | 2020-09-08 16:06:32 +0100 | [diff] [blame] | 3342 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model, |
| 3343 | data, nullptr, validateFunc, activation); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3344 | } |
| 3345 | |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3346 | template<typename HalPolicy, |
| 3347 | typename HalOperation = typename HalPolicy::Operation, |
| 3348 | typename HalModel = typename HalPolicy::Model> |
| 3349 | bool ConvertUnidirectionalSequenceLstm(const HalOperation& operation, |
| 3350 | const HalModel& model, |
| 3351 | ConversionData& data) |
| 3352 | { |
| 3353 | using HalOperand = typename HalPolicy::Operand; |
| 3354 | using HalOperandType = typename HalPolicy::OperandType; |
| 3355 | |
| 3356 | ALOGV("HalPolicy::ConvertUnidirectionalSequenceLstm()"); |
| 3357 | |
| 3358 | // Determine if input OperandType is ANEURALNETWORKS_TENSOR_FLOAT 32 or 16 |
| 3359 | HalOperandType inputType; |
| 3360 | if (!GetOperandType<HalPolicy>(operation, 0, model, inputType)) |
| 3361 | { |
| 3362 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3363 | } |
| 3364 | |
| 3365 | // Inputs: |
| 3366 | // 0: The input: A 3-D tensor of shape: If time-major: [max_time, batch_size, input_size] If batch-major: |
| 3367 | // [batch_size, max_time, input_size] where “max_time” is the number of timesteps (sequence length), “batch_size” |
| 3368 | // corresponds to the batching dimension, and “input_size” is the size of the input. |
| 3369 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 3370 | if (!input.IsValid()) |
| 3371 | { |
| 3372 | return Fail("%s: Could not read input 0: input", __func__); |
| 3373 | } |
| 3374 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [batch_size, output_size]. |
| 3375 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 18, model, data); |
| 3376 | if (!outputStateIn.IsValid()) |
| 3377 | { |
| 3378 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 3379 | } |
| 3380 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [batch_size, num_units]. |
| 3381 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle<HalPolicy>(operation, 19, model, data); |
| 3382 | if (!cellStateIn.IsValid()) |
| 3383 | { |
| 3384 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 3385 | } |
| 3386 | |
| 3387 | // Get the mandatory input tensors: |
| 3388 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3389 | // [num_units, input_size]. |
| 3390 | const ConstTensorPin inputToForgetWeightsPin = |
| 3391 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 2)); |
| 3392 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3393 | // [num_units, input_size]. |
| 3394 | const ConstTensorPin inputToCellWeightsPin = |
| 3395 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 3)); |
| 3396 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3397 | // [num_units, input_size]. |
| 3398 | const ConstTensorPin inputToOutputWeightsPin = |
| 3399 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 4)); |
| 3400 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3401 | // [num_units, output_size]. |
| 3402 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 3403 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 6)); |
| 3404 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 3405 | // [num_units, output_size]. |
| 3406 | const ConstTensorPin recurrentToCellWeightsPin = |
| 3407 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 7)); |
| 3408 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3409 | // [num_units, output_size]. |
| 3410 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 3411 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 8)); |
| 3412 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3413 | const ConstTensorPin forgetGateBiasPin = |
| 3414 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 13, model, data); |
| 3415 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3416 | const ConstTensorPin cellBiasPin = |
| 3417 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 14, model, data); |
| 3418 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3419 | const ConstTensorPin outputGateBiasPin = |
| 3420 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 15, model, data); |
| 3421 | |
| 3422 | if (!inputToForgetWeightsPin.IsValid() || |
| 3423 | !inputToCellWeightsPin.IsValid() || |
| 3424 | !inputToOutputWeightsPin.IsValid() || |
| 3425 | !recurrentToForgetWeightsPin.IsValid() || |
| 3426 | !recurrentToCellWeightsPin.IsValid() || |
| 3427 | !recurrentToOutputWeightsPin.IsValid() || |
| 3428 | !forgetGateBiasPin.IsValid() || |
| 3429 | !cellBiasPin.IsValid() || |
| 3430 | !outputGateBiasPin.IsValid()) |
| 3431 | { |
| 3432 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3433 | } |
| 3434 | |
| 3435 | // Get the optional input tensors: |
| 3436 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3437 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 3438 | const ConstTensorPin inputToInputWeightsPin = |
| 3439 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 1, true)); |
| 3440 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3441 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 3442 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 3443 | const ConstTensorPin recurrentToInputWeightsPin = |
| 3444 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 5, true)); |
| 3445 | // 09: The cell-to-input weights: Optional. |
| 3446 | // A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3447 | const ConstTensorPin cellToInputWeightsPin = |
| 3448 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 9, true)); |
| 3449 | // 10: The cell-to-forget weights: Optional. |
| 3450 | // A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3451 | const ConstTensorPin cellToForgetWeightsPin = |
| 3452 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 10, true)); |
| 3453 | // 11: The cell-to-output weights: Optional. |
| 3454 | // A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3455 | const ConstTensorPin cellToOutputWeightsPin = |
| 3456 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 11, true)); |
| 3457 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [num_units]. |
| 3458 | const ConstTensorPin inputGateBiasPin = |
| 3459 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3460 | 12, |
| 3461 | model, |
| 3462 | data, |
| 3463 | g_DontPermute, |
| 3464 | nullptr, |
| 3465 | true); |
| 3466 | |
| 3467 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape |
| 3468 | // [output_size, num_units]. |
| 3469 | const ConstTensorPin projectionWeightsPin = |
| 3470 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 16, true)); |
| 3471 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16, of shape [output_size]. |
| 3472 | const ConstTensorPin projectionBiasPin = |
| 3473 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3474 | 17, |
| 3475 | model, |
| 3476 | data, |
| 3477 | g_DontPermute, |
| 3478 | nullptr, |
| 3479 | true); |
| 3480 | |
| 3481 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 3482 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 3483 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 3484 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 3485 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 3486 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 3487 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 3488 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 3489 | { |
| 3490 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3491 | } |
| 3492 | |
| 3493 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 3494 | // 20: The activation function: A value indicating the activation function: |
| 3495 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 3496 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 3497 | // If set to 0.0 then clipping is disabled. |
| 3498 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 3499 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 3500 | // Determine data type of input tensor |
Cathal Corbett | 3fe3a54 | 2022-05-25 09:35:37 +0100 | [diff] [blame] | 3501 | ActivationFn activation = ActivationFn::kActivationNone; |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3502 | LstmDescriptor desc; |
| 3503 | |
| 3504 | if (inputType == HalOperandType::TENSOR_FLOAT32) |
| 3505 | { |
| 3506 | float cellClip; |
| 3507 | float projClip; |
| 3508 | |
| 3509 | if (!GetInputActivationFunctionFromTensor<HalPolicy>(operation, 20, activation, model, data) || |
| 3510 | !GetInputScalar<HalPolicy>(operation, 21, HalOperandType::FLOAT32, cellClip, model, data) || |
| 3511 | !GetInputScalar<HalPolicy>(operation, 22, HalOperandType::FLOAT32, projClip, model, data)) |
| 3512 | { |
| 3513 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 3514 | } |
| 3515 | |
| 3516 | desc.m_ClippingThresCell = cellClip; |
| 3517 | desc.m_ClippingThresProj = projClip; |
| 3518 | } |
| 3519 | |
| 3520 | if (inputType == HalOperandType::TENSOR_FLOAT16) |
| 3521 | { |
| 3522 | Half cellClip; |
| 3523 | Half projClip; |
| 3524 | |
| 3525 | if (!GetInputActivationFunctionFromTensor<HalPolicy>(operation, 20, activation, model, data) || |
| 3526 | !GetInputScalar<HalPolicy>(operation, 21, HalOperandType::FLOAT16, cellClip, model, data) || |
| 3527 | !GetInputScalar<HalPolicy>(operation, 22, HalOperandType::FLOAT16, projClip, model, data)) |
| 3528 | { |
| 3529 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 3530 | } |
| 3531 | |
| 3532 | desc.m_ClippingThresCell = cellClip; |
| 3533 | desc.m_ClippingThresProj = projClip; |
| 3534 | } |
| 3535 | |
| 3536 | // Determine if time-major or batch-major. |
| 3537 | // 23: Time-major if true, batch-major if false. |
| 3538 | bool isTimeMajor = GetOptionalBool<HalPolicy>(operation, 23, model, data); |
| 3539 | |
| 3540 | // Get the normalization tensors |
| 3541 | // 24: The input layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3542 | // Used to rescale normalized inputs to activation at input gate. |
| 3543 | const ConstTensorPin inputLayerNormWeightsPin |
| 3544 | (DequantizeAndMakeConstTensorPin<HalPolicy>(operation, model, data, 24, true)); |
| 3545 | |
| 3546 | // 25: The forget layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3547 | // Used to rescale normalized inputs to activation at forget gate. |
| 3548 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 3549 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3550 | 25, |
| 3551 | model, |
| 3552 | data, |
| 3553 | g_DontPermute, |
| 3554 | nullptr, |
| 3555 | true); |
| 3556 | |
| 3557 | // 26: The cell layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3558 | // Used to rescale normalized inputs to activation at cell gate. |
| 3559 | const ConstTensorPin cellLayerNormWeightsPin = |
| 3560 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3561 | 26, |
| 3562 | model, |
| 3563 | data, |
| 3564 | g_DontPermute, |
| 3565 | nullptr, |
| 3566 | true); |
| 3567 | |
| 3568 | // 27: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3569 | // Used to rescale normalized inputs to activation at output gate. |
| 3570 | const ConstTensorPin outputLayerNormWeightsPin = |
| 3571 | ConvertOperationInputToConstTensorPin<HalPolicy>(operation, |
| 3572 | 27, |
| 3573 | model, |
| 3574 | data, |
| 3575 | g_DontPermute, |
| 3576 | nullptr, |
| 3577 | true); |
| 3578 | |
| 3579 | // Outputs: |
| 3580 | // 00: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32/16. Shape: if time-major: |
| 3581 | // [max_time, batch_size, output_size] If batch-major: [batch_size, max_time, output_size] |
| 3582 | const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 3583 | if (!output) |
| 3584 | { |
| 3585 | return Fail("%s: Could not read output: ", __func__); |
| 3586 | } |
| 3587 | |
| 3588 | // |
| 3589 | // 01 & 02: |
| 3590 | // hiddenStateOut and cellStateOut are not currently supported by our android versioning. |
| 3591 | // |
| 3592 | |
| 3593 | // set the params structure for the AddLstmLayer call |
| 3594 | LstmInputParams params; |
| 3595 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 3596 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 3597 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 3598 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 3599 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 3600 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 3601 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 3602 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 3603 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 3604 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 3605 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 3606 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 3607 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 3608 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 3609 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 3610 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 3611 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 3612 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 3613 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 3614 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 3615 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 3616 | |
| 3617 | // set the layer descriptor |
| 3618 | desc.m_ActivationFunc = activation; |
| 3619 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 3620 | params.m_RecurrentToInputWeights == nullptr || |
| 3621 | params.m_InputGateBias == nullptr); |
| 3622 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 3623 | params.m_CellToOutputWeights != nullptr); |
| 3624 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 3625 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 3626 | params.m_ForgetLayerNormWeights != nullptr || |
| 3627 | params.m_CellLayerNormWeights != nullptr || |
| 3628 | params.m_OutputLayerNormWeights != nullptr); |
| 3629 | desc.m_TimeMajor = isTimeMajor; |
| 3630 | |
| 3631 | // validate the optional input groups |
| 3632 | if (desc.m_CifgEnabled && |
| 3633 | (params.m_InputToInputWeights != nullptr || |
| 3634 | params.m_RecurrentToInputWeights != nullptr || |
| 3635 | params.m_InputGateBias != nullptr)) |
| 3636 | { |
| 3637 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 3638 | " and input gate bias must be provided", __func__); |
| 3639 | } |
| 3640 | |
| 3641 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 3642 | { |
| 3643 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 3644 | } |
| 3645 | |
| 3646 | if (desc.m_PeepholeEnabled && |
| 3647 | (params.m_CellToForgetWeights == nullptr || |
| 3648 | params.m_CellToOutputWeights == nullptr || |
| 3649 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 3650 | { |
| 3651 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 3652 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 3653 | } |
| 3654 | |
| 3655 | if (desc.m_LayerNormEnabled && |
| 3656 | (params.m_ForgetLayerNormWeights == nullptr || |
| 3657 | params.m_CellLayerNormWeights == nullptr || |
| 3658 | params.m_OutputLayerNormWeights == nullptr || |
| 3659 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 3660 | { |
| 3661 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 3662 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 3663 | } |
| 3664 | |
| 3665 | // Check if the layer is supported |
| 3666 | // Inputs |
| 3667 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3668 | const TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 3669 | const TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
| 3670 | |
| 3671 | // Outputs |
| 3672 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3673 | |
Mike Kelly | 0ae102a | 2022-04-25 16:18:57 +0100 | [diff] [blame] | 3674 | unsigned int batchSize = inputInfo.GetShape()[0]; |
| 3675 | unsigned int outputSize = outputInfo.GetShape()[2]; |
| 3676 | unsigned int numUnits = cellStateInInfo.GetShape()[1]; |
| 3677 | |
| 3678 | armnn::DataType dataType = inputInfo.GetDataType(); |
| 3679 | float qScale = inputInfo.GetQuantizationScale(); |
| 3680 | int qOffset = inputInfo.GetQuantizationOffset(); |
| 3681 | |
| 3682 | armnn::TensorInfo cellStateOutInfo({batchSize, numUnits}, cellStateInInfo.GetDataType(), |
| 3683 | cellStateInInfo.GetQuantizationScale(), cellStateInInfo.GetQuantizationOffset()); |
| 3684 | armnn::TensorInfo outputStateOutInfo({batchSize, outputSize}, dataType, qScale, qOffset); |
| 3685 | |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3686 | // Basic parameters |
| 3687 | LstmInputParamsInfo paramsInfo; |
| 3688 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 3689 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 3690 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 3691 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 3692 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 3693 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 3694 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 3695 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 3696 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 3697 | |
| 3698 | // Optional parameters |
| 3699 | if (!desc.m_CifgEnabled) |
| 3700 | { |
| 3701 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 3702 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 3703 | if (params.m_CellToInputWeights != nullptr) |
| 3704 | { |
| 3705 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 3706 | } |
| 3707 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 3708 | } |
| 3709 | |
| 3710 | if (desc.m_ProjectionEnabled) |
| 3711 | { |
| 3712 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 3713 | if (params.m_ProjectionBias != nullptr) |
| 3714 | { |
| 3715 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 3716 | } |
| 3717 | } |
| 3718 | |
| 3719 | if (desc.m_PeepholeEnabled) |
| 3720 | { |
| 3721 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 3722 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 3723 | } |
| 3724 | |
| 3725 | if (desc.m_LayerNormEnabled) |
| 3726 | { |
| 3727 | if(!desc.m_CifgEnabled) |
| 3728 | { |
| 3729 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 3730 | } |
| 3731 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 3732 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 3733 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 3734 | } |
| 3735 | |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3736 | bool isSupported = false; |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3737 | armnn::BackendId setBackend; |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3738 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3739 | { |
| 3740 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3741 | IsUnidirectionalSequenceLstmSupported, |
| 3742 | data.m_Backends, |
| 3743 | isSupported, |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3744 | setBackend, |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3745 | inputInfo, |
| 3746 | outputStateInInfo, |
| 3747 | cellStateInInfo, |
Mike Kelly | 0ae102a | 2022-04-25 16:18:57 +0100 | [diff] [blame] | 3748 | outputStateOutInfo, |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3749 | cellStateOutInfo, |
Mike Kelly | 0ae102a | 2022-04-25 16:18:57 +0100 | [diff] [blame] | 3750 | outputInfo, |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3751 | desc, |
| 3752 | paramsInfo); |
| 3753 | }; |
| 3754 | |
| 3755 | bool isDynamic = false; |
| 3756 | if (!IsDynamicTensor(outputInfo)) |
| 3757 | { |
| 3758 | validateFunc(outputInfo, isSupported); |
| 3759 | } |
| 3760 | else |
| 3761 | { |
| 3762 | isDynamic = true; |
| 3763 | isSupported = AreDynamicTensorsSupported(); |
| 3764 | } |
| 3765 | |
| 3766 | if (!isSupported) |
| 3767 | { |
| 3768 | return false; |
| 3769 | } |
| 3770 | |
| 3771 | // Add the layer |
| 3772 | IConnectableLayer* layer = data.m_Network->AddUnidirectionalSequenceLstmLayer(desc, |
| 3773 | params, |
| 3774 | "UnidirectionalSequenceLstm"); |
Cathal Corbett | 8de96f7 | 2022-09-01 13:34:59 +0100 | [diff] [blame] | 3775 | layer->SetBackendId(setBackend); |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3776 | |
| 3777 | input.Connect(layer->GetInputSlot(0)); |
| 3778 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 3779 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 3780 | |
| 3781 | if (!isDynamic) |
| 3782 | { |
Mike Kelly | 0ae102a | 2022-04-25 16:18:57 +0100 | [diff] [blame] | 3783 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 2, model, data)); |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3784 | } |
| 3785 | else |
| 3786 | { |
Mike Kelly | 0ae102a | 2022-04-25 16:18:57 +0100 | [diff] [blame] | 3787 | return (SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 2, model, data, nullptr, |
| 3788 | validateFunc, ActivationFn::kActivationNone, true)); |
Cathal Corbett | 0fa5e6d | 2022-01-21 16:55:13 +0000 | [diff] [blame] | 3789 | } |
| 3790 | } |
| 3791 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 3792 | } // armnn_driver namespace |