Laurent Carlier | 749294b | 2020-06-01 09:03:17 +0100 | [diff] [blame] | 1 | // |
Mike Kelly | 3ec3077 | 2023-03-08 13:47:17 +0000 | [diff] [blame] | 2 | // Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 5 | |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 6 | #include "NeonBackendId.hpp" |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 7 | #include "NeonBackendModelContext.hpp" |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 8 | #include "NeonTensorHandle.hpp" |
| 9 | #include "NeonWorkloadFactory.hpp" |
| 10 | |
David Beck | 0dbe0ee | 2018-09-24 15:59:27 +0100 | [diff] [blame] | 11 | #include <Layer.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 12 | |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 13 | #include <armnn/Utils.hpp> |
Jan Eilers | 8eb2560 | 2020-03-09 12:13:48 +0000 | [diff] [blame] | 14 | #include <armnn/utility/IgnoreUnused.hpp> |
Matthew Sloyan | 171214c | 2020-09-09 09:07:37 +0100 | [diff] [blame] | 15 | #include <armnn/utility/NumericCast.hpp> |
Jan Eilers | bb446e5 | 2020-04-02 13:56:54 +0100 | [diff] [blame] | 16 | #include <armnn/utility/PolymorphicDowncast.hpp> |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 17 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 18 | #include <backendsCommon/MakeWorkloadHelper.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 19 | #include <armnn/backends/MemCopyWorkload.hpp> |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 20 | #include <backendsCommon/MemImportWorkload.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 21 | #include <armnn/backends/TensorHandle.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 22 | |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 23 | #include <neon/workloads/NeonWorkloadUtils.hpp> |
| 24 | #include <neon/workloads/NeonWorkloads.hpp> |
| 25 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 26 | namespace armnn |
| 27 | { |
| 28 | |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 29 | namespace |
| 30 | { |
| 31 | static const BackendId s_Id{NeonBackendId()}; |
| 32 | } |
| 33 | |
David Beck | 29c75de | 2018-10-23 13:35:58 +0100 | [diff] [blame] | 34 | bool NeonWorkloadFactory::IsLayerSupported(const Layer& layer, |
| 35 | Optional<DataType> dataType, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 36 | std::string& outReasonIfUnsupported) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 37 | { |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 38 | return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported); |
| 39 | } |
| 40 | |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 41 | bool NeonWorkloadFactory::IsLayerSupported(const IConnectableLayer& layer, |
| 42 | Optional<DataType> dataType, |
| 43 | std::string& outReasonIfUnsupported, |
| 44 | const ModelOptions& modelOptions) |
| 45 | { |
| 46 | return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported, modelOptions); |
| 47 | } |
| 48 | |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 49 | const BackendId& NeonWorkloadFactory::GetBackendId() const |
| 50 | { |
| 51 | return s_Id; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 52 | } |
| 53 | |
Matthew Sloyan | 0a7dc6b | 2021-02-10 16:50:53 +0000 | [diff] [blame] | 54 | void NeonWorkloadFactory::SetNumberOfThreads() |
| 55 | { |
| 56 | if (m_ModelContextPtr) |
| 57 | { |
| 58 | const unsigned int MIN_THREADS = 1; |
| 59 | const unsigned int MAX_THREADS = 64; |
| 60 | |
| 61 | // Set the number of threads to be used if the user has set NumberOfThreads param |
| 62 | // Only set if within limit or valid input |
| 63 | auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get()); |
| 64 | auto numberOfThreads = modelOptions->GetNumberOfThreads(); |
| 65 | |
| 66 | if (numberOfThreads != 0 && numberOfThreads >= MIN_THREADS && numberOfThreads <= MAX_THREADS) |
| 67 | { |
| 68 | arm_compute::Scheduler::get().set_num_threads(numberOfThreads); |
| 69 | } |
| 70 | } |
| 71 | } |
| 72 | |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 73 | NeonWorkloadFactory::NeonWorkloadFactory(const std::shared_ptr<NeonMemoryManager>& memoryManager) |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 74 | : m_MemoryManager(memoryManager), m_ModelContextPtr(IBackendInternal::IBackendSpecificModelContextPtr{}) |
| 75 | { |
Matthew Sloyan | 0a7dc6b | 2021-02-10 16:50:53 +0000 | [diff] [blame] | 76 | SetNumberOfThreads(); |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 77 | } |
| 78 | |
| 79 | NeonWorkloadFactory::NeonWorkloadFactory(const std::shared_ptr<NeonMemoryManager>& memoryManager, |
| 80 | const IBackendInternal::IBackendSpecificModelContextPtr& modelContextPtr) |
| 81 | : m_MemoryManager(memoryManager), m_ModelContextPtr(modelContextPtr) |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 82 | { |
Matthew Sloyan | 0a7dc6b | 2021-02-10 16:50:53 +0000 | [diff] [blame] | 83 | SetNumberOfThreads(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 84 | } |
| 85 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 86 | std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateSubTensorHandle(ITensorHandle& parent, |
| 87 | TensorShape const& subTensorShape, |
| 88 | unsigned int const* subTensorOrigin) const |
| 89 | { |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 90 | const arm_compute::TensorShape shape = armcomputetensorutils::BuildArmComputeTensorShape(subTensorShape); |
| 91 | |
| 92 | arm_compute::Coordinates coords; |
| 93 | coords.set_num_dimensions(subTensorShape.GetNumDimensions()); |
| 94 | for (unsigned int i = 0; i < subTensorShape.GetNumDimensions(); i++) |
| 95 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 96 | // Arm compute indexes tensor coords in reverse order. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 97 | unsigned int revertedIndex = subTensorShape.GetNumDimensions() - i - 1; |
Matthew Sloyan | 171214c | 2020-09-09 09:07:37 +0100 | [diff] [blame] | 98 | coords.set(i, armnn::numeric_cast<int>(subTensorOrigin[revertedIndex])); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 99 | } |
| 100 | |
Derek Lamberti | 0790dce | 2019-04-15 18:37:35 +0100 | [diff] [blame] | 101 | const arm_compute::TensorShape parentShape = armcomputetensorutils::BuildArmComputeTensorShape(parent.GetShape()); |
| 102 | if (!::arm_compute::error_on_invalid_subtensor(__func__, __FILE__, __LINE__, parentShape, coords, shape)) |
| 103 | { |
| 104 | return nullptr; |
| 105 | } |
| 106 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 107 | return std::make_unique<NeonSubTensorHandle>( |
Jan Eilers | bb446e5 | 2020-04-02 13:56:54 +0100 | [diff] [blame] | 108 | PolymorphicDowncast<IAclTensorHandle*>(&parent), shape, coords); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 109 | } |
| 110 | |
David Monahan | 3fb7e10 | 2019-08-20 11:25:29 +0100 | [diff] [blame] | 111 | std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo, |
| 112 | const bool IsMemoryManaged) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 113 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 114 | auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo); |
David Monahan | 3fb7e10 | 2019-08-20 11:25:29 +0100 | [diff] [blame] | 115 | if (IsMemoryManaged) |
| 116 | { |
| 117 | tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup()); |
| 118 | } |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 119 | return tensorHandle; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 120 | } |
| 121 | |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 122 | std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo, |
David Monahan | 3fb7e10 | 2019-08-20 11:25:29 +0100 | [diff] [blame] | 123 | DataLayout dataLayout, |
| 124 | const bool IsMemoryManaged) const |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 125 | { |
| 126 | auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo, dataLayout); |
David Monahan | 3fb7e10 | 2019-08-20 11:25:29 +0100 | [diff] [blame] | 127 | if (IsMemoryManaged) |
| 128 | { |
| 129 | tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup()); |
| 130 | } |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 131 | return tensorHandle; |
| 132 | } |
| 133 | |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 134 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateWorkload(LayerType type, |
| 135 | const QueueDescriptor& descriptor, |
| 136 | const WorkloadInfo& info) const |
| 137 | { |
| 138 | switch(type) |
| 139 | { |
| 140 | case LayerType::Activation : |
| 141 | { |
| 142 | auto activationQueueDescriptor = PolymorphicDowncast<const ActivationQueueDescriptor*>(&descriptor); |
| 143 | return std::make_unique<NeonActivationWorkload>(*activationQueueDescriptor, info); |
| 144 | } |
| 145 | case LayerType::Addition : |
| 146 | { |
| 147 | auto additionQueueDescriptor = PolymorphicDowncast<const AdditionQueueDescriptor*>(&descriptor); |
| 148 | return std::make_unique<NeonAdditionWorkload>(*additionQueueDescriptor, info); |
| 149 | } |
| 150 | case LayerType::ArgMinMax : |
| 151 | { |
| 152 | auto argMinMaxQueueDescriptor = PolymorphicDowncast<const ArgMinMaxQueueDescriptor*>(&descriptor); |
| 153 | return std::make_unique<NeonArgMinMaxWorkload>(*argMinMaxQueueDescriptor, info); |
| 154 | } |
Teresa Charlin | 0f86ecf | 2022-10-13 15:47:08 +0100 | [diff] [blame] | 155 | case LayerType::BatchMatMul : |
| 156 | { |
| 157 | auto batchMatMulQueueDescriptor = PolymorphicDowncast<const BatchMatMulQueueDescriptor*>(&descriptor); |
Teresa Charlin | 1fe6c81 | 2022-11-01 15:59:50 +0000 | [diff] [blame] | 158 | bool isFastMathEnabled = false; |
| 159 | if (m_ModelContextPtr) |
| 160 | { |
| 161 | if (m_ModelContextPtr.get() != nullptr) |
| 162 | { |
| 163 | auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get()); |
| 164 | if (modelOptions) |
| 165 | { |
| 166 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 167 | } |
| 168 | } |
| 169 | } |
| 170 | return std::make_unique<NeonBatchMatMulWorkload>(*batchMatMulQueueDescriptor, info, isFastMathEnabled); |
Teresa Charlin | 0f86ecf | 2022-10-13 15:47:08 +0100 | [diff] [blame] | 171 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 172 | case LayerType::BatchNormalization : |
| 173 | { |
| 174 | auto batchNormalizationQueueDescriptor |
| 175 | = PolymorphicDowncast<const BatchNormalizationQueueDescriptor*>(&descriptor); |
| 176 | return std::make_unique<NeonBatchNormalizationWorkload>(*batchNormalizationQueueDescriptor, info); |
| 177 | } |
| 178 | case LayerType::BatchToSpaceNd : |
| 179 | { |
| 180 | auto batchToSpaceNdQueueDescriptor |
| 181 | = PolymorphicDowncast<const BatchToSpaceNdQueueDescriptor*>(&descriptor); |
| 182 | return std::make_unique<NeonBatchToSpaceNdWorkload>(*batchToSpaceNdQueueDescriptor, info); |
| 183 | } |
| 184 | case LayerType::Cast : |
| 185 | { |
| 186 | auto castQueueDescriptor = PolymorphicDowncast<const CastQueueDescriptor*>(&descriptor); |
| 187 | return std::make_unique<NeonCastWorkload>(*castQueueDescriptor, info); |
| 188 | } |
| 189 | case LayerType::ChannelShuffle : |
| 190 | { |
| 191 | auto channelShuffleQueueDescriptor = PolymorphicDowncast<const ChannelShuffleQueueDescriptor*>(&descriptor); |
| 192 | return std::make_unique<NeonChannelShuffleWorkload>(*channelShuffleQueueDescriptor, info); |
| 193 | } |
| 194 | case LayerType::Comparison : |
| 195 | { |
| 196 | auto comparisonQueueDescriptor = PolymorphicDowncast<const ComparisonQueueDescriptor*>(&descriptor); |
| 197 | return std::make_unique<NeonComparisonWorkload>(*comparisonQueueDescriptor, info); |
| 198 | } |
| 199 | case LayerType::Concat : |
| 200 | { |
| 201 | auto concatQueueDescriptor = PolymorphicDowncast<const ConcatQueueDescriptor*>(&descriptor); |
| 202 | return std::make_unique<NeonConcatWorkload>(*concatQueueDescriptor, info); |
| 203 | } |
| 204 | case LayerType::Constant : |
| 205 | { |
| 206 | auto constantQueueDescriptor = PolymorphicDowncast<const ConstantQueueDescriptor*>(&descriptor); |
| 207 | return std::make_unique<NeonConstantWorkload>(*constantQueueDescriptor, info); |
| 208 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 209 | case LayerType::ConvertFp16ToFp32 : |
| 210 | { |
| 211 | auto convertFp16ToFp32QueueDescriptor |
| 212 | = PolymorphicDowncast<const ConvertFp16ToFp32QueueDescriptor*>(&descriptor); |
| 213 | return std::make_unique<NeonConvertFp16ToFp32Workload>(*convertFp16ToFp32QueueDescriptor, info); |
| 214 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 215 | case LayerType::ConvertFp32ToFp16 : |
| 216 | { |
| 217 | auto convertFp32ToFp16QueueDescriptor |
| 218 | = PolymorphicDowncast<const ConvertFp32ToFp16QueueDescriptor*>(&descriptor); |
| 219 | return std::make_unique<NeonConvertFp32ToFp16Workload>(*convertFp32ToFp16QueueDescriptor, info); |
| 220 | } |
| 221 | case LayerType::Convolution2d : |
| 222 | { |
| 223 | auto convolution2dQueueDescriptor = PolymorphicDowncast<const Convolution2dQueueDescriptor*>(&descriptor); |
| 224 | |
| 225 | bool isFastMathEnabled = false; |
| 226 | if (m_ModelContextPtr) |
| 227 | { |
| 228 | if (m_ModelContextPtr.get() != nullptr) |
| 229 | { |
| 230 | auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get()); |
| 231 | if (modelOptions) |
| 232 | { |
| 233 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 234 | } |
| 235 | } |
| 236 | } |
| 237 | return std::make_unique<NeonConvolution2dWorkload>(*convolution2dQueueDescriptor, |
| 238 | info, |
| 239 | m_MemoryManager->GetIntraLayerManager(), |
| 240 | isFastMathEnabled); |
| 241 | } |
| 242 | case LayerType::Convolution3d : |
| 243 | { |
| 244 | auto convolution3dQueueDescriptor = PolymorphicDowncast<const Convolution3dQueueDescriptor*>(&descriptor); |
| 245 | |
| 246 | bool isFastMathEnabled = false; |
| 247 | if (m_ModelContextPtr) |
| 248 | { |
| 249 | if (m_ModelContextPtr.get() != nullptr) |
| 250 | { |
| 251 | auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get()); |
| 252 | if (modelOptions) |
| 253 | { |
| 254 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 255 | } |
| 256 | } |
| 257 | } |
| 258 | return std::make_unique<NeonConvolution3dWorkload>(*convolution3dQueueDescriptor, |
| 259 | info, |
| 260 | m_MemoryManager->GetIntraLayerManager(), |
| 261 | isFastMathEnabled); |
| 262 | } |
| 263 | case LayerType::Debug : |
| 264 | { |
| 265 | auto debugQueueDescriptor = PolymorphicDowncast<const DebugQueueDescriptor*>(&descriptor); |
| 266 | return MakeWorkloadHelper<NullWorkload, NullWorkload>(*debugQueueDescriptor, info); |
| 267 | } |
| 268 | case LayerType::DepthToSpace : |
| 269 | { |
| 270 | auto depthToSpaceQueueDescriptor = PolymorphicDowncast<const DepthToSpaceQueueDescriptor*>(&descriptor); |
| 271 | return std::make_unique<NeonDepthToSpaceWorkload>(*depthToSpaceQueueDescriptor, info); |
| 272 | } |
| 273 | case LayerType::DepthwiseConvolution2d : |
| 274 | { |
| 275 | auto depthwiseConvolution2dQueueDescriptor |
| 276 | = PolymorphicDowncast<const DepthwiseConvolution2dQueueDescriptor*>(&descriptor); |
| 277 | return std::make_unique<NeonDepthwiseConvolutionWorkload>(*depthwiseConvolution2dQueueDescriptor, info); |
| 278 | } |
| 279 | case LayerType::Dequantize : |
| 280 | { |
| 281 | auto dequantizeQueueDescriptor = PolymorphicDowncast<const DequantizeQueueDescriptor*>(&descriptor); |
| 282 | return std::make_unique<NeonDequantizeWorkload>(*dequantizeQueueDescriptor, info); |
| 283 | } |
| 284 | case LayerType::DetectionPostProcess : |
| 285 | { |
| 286 | auto detectionPostProcessQueueDescriptor |
| 287 | = PolymorphicDowncast<const DetectionPostProcessQueueDescriptor*>(&descriptor); |
| 288 | return MakeWorkloadHelper<NullWorkload, NullWorkload>(*detectionPostProcessQueueDescriptor, info); |
| 289 | } |
| 290 | case LayerType::Division : |
| 291 | { |
| 292 | auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor); |
| 293 | return std::make_unique<NeonDivisionWorkload>(*divisionQueueDescriptor, info); |
| 294 | } |
Mike Kelly | 3ec3077 | 2023-03-08 13:47:17 +0000 | [diff] [blame] | 295 | case LayerType::ElementwiseBinary : |
| 296 | { |
| 297 | auto elementwiseBinaryQueueDescriptor |
| 298 | = PolymorphicDowncast<const ElementwiseBinaryQueueDescriptor*>(&descriptor); |
| 299 | |
| 300 | switch (elementwiseBinaryQueueDescriptor->m_Parameters.m_Operation) |
| 301 | { |
| 302 | case BinaryOperation::Add: |
| 303 | { |
| 304 | AdditionQueueDescriptor additionQueueDescriptor; |
| 305 | additionQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 306 | additionQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
| 307 | return std::make_unique<NeonAdditionWorkload>(additionQueueDescriptor, info); |
| 308 | } |
| 309 | case BinaryOperation::Div: |
| 310 | { |
| 311 | DivisionQueueDescriptor divisionQueueDescriptor; |
| 312 | divisionQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 313 | divisionQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
| 314 | return std::make_unique<NeonDivisionWorkload>(divisionQueueDescriptor, info); |
| 315 | } |
| 316 | case BinaryOperation::Maximum: |
| 317 | { |
| 318 | MaximumQueueDescriptor maximumQueueDescriptor; |
| 319 | maximumQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 320 | maximumQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
| 321 | return std::make_unique<NeonMaximumWorkload>(maximumQueueDescriptor, info); |
| 322 | } |
| 323 | case BinaryOperation::Minimum: |
| 324 | { |
| 325 | MinimumQueueDescriptor minimumQueueDescriptor; |
| 326 | minimumQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 327 | minimumQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
| 328 | return std::make_unique<NeonMinimumWorkload>(minimumQueueDescriptor, info); |
| 329 | } |
| 330 | case BinaryOperation::Mul: |
| 331 | { |
| 332 | MultiplicationQueueDescriptor multiplicationQueueDescriptor; |
| 333 | multiplicationQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 334 | multiplicationQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
| 335 | return std::make_unique<NeonMultiplicationWorkload>(multiplicationQueueDescriptor, info); |
| 336 | } |
John Mcloughlin | 34c1c38 | 2023-05-17 15:08:36 +0100 | [diff] [blame] | 337 | case BinaryOperation::Power: |
| 338 | case BinaryOperation::SqDiff: |
| 339 | { |
| 340 | return std::make_unique<NeonElementwiseBinaryWorkload>(*elementwiseBinaryQueueDescriptor, info); |
| 341 | } |
Mike Kelly | 3ec3077 | 2023-03-08 13:47:17 +0000 | [diff] [blame] | 342 | case BinaryOperation::Sub: |
| 343 | { |
| 344 | SubtractionQueueDescriptor subtractionQueueDescriptor; |
| 345 | subtractionQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 346 | subtractionQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
| 347 | return std::make_unique<NeonSubtractionWorkload>(subtractionQueueDescriptor, info); |
| 348 | } |
| 349 | default: |
| 350 | return nullptr; |
| 351 | } |
| 352 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 353 | case LayerType::ElementwiseUnary : |
| 354 | { |
| 355 | auto elementwiseUnaryQueueDescriptor |
| 356 | = PolymorphicDowncast<const ElementwiseUnaryQueueDescriptor*>(&descriptor); |
| 357 | |
| 358 | switch(elementwiseUnaryQueueDescriptor->m_Parameters.m_Operation) |
| 359 | { |
| 360 | case UnaryOperation::Abs: |
| 361 | { |
| 362 | AbsQueueDescriptor absQueueDescriptor; |
| 363 | absQueueDescriptor.m_Inputs = elementwiseUnaryQueueDescriptor->m_Inputs; |
| 364 | absQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs; |
| 365 | |
| 366 | return std::make_unique<NeonAbsWorkload>(absQueueDescriptor, info); |
| 367 | } |
| 368 | case UnaryOperation::Exp: |
| 369 | return std::make_unique<NeonExpWorkload>(*elementwiseUnaryQueueDescriptor, info); |
| 370 | case UnaryOperation::LogicalNot: |
| 371 | return std::make_unique<NeonLogicalNotWorkload>(*elementwiseUnaryQueueDescriptor, info); |
| 372 | case UnaryOperation::Log: |
| 373 | return std::make_unique<NeonLogWorkload>(*elementwiseUnaryQueueDescriptor, info); |
| 374 | case UnaryOperation::Neg: |
| 375 | return std::make_unique<NeonNegWorkload>(*elementwiseUnaryQueueDescriptor, info); |
| 376 | case UnaryOperation::Rsqrt: |
| 377 | { |
| 378 | RsqrtQueueDescriptor rsqrtQueueDescriptor; |
| 379 | rsqrtQueueDescriptor.m_Inputs = elementwiseUnaryQueueDescriptor->m_Inputs; |
| 380 | rsqrtQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs; |
| 381 | |
| 382 | return std::make_unique<NeonRsqrtWorkload>(rsqrtQueueDescriptor, info); |
| 383 | } |
| 384 | case UnaryOperation::Sin: |
| 385 | return std::make_unique<NeonSinWorkload>(*elementwiseUnaryQueueDescriptor, info); |
Teresa Charlin | 0302723 | 2022-05-09 17:27:08 +0100 | [diff] [blame] | 386 | case UnaryOperation::Sqrt: |
| 387 | return std::make_unique<NeonSqrtWorkload>(*elementwiseUnaryQueueDescriptor, info); |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 388 | default: |
| 389 | return nullptr; |
| 390 | } |
| 391 | } |
| 392 | case LayerType::Fill : |
| 393 | { |
| 394 | auto fillQueueDescriptor = PolymorphicDowncast<const FillQueueDescriptor*>(&descriptor); |
| 395 | return std::make_unique<NeonFillWorkload>(*fillQueueDescriptor, info); |
| 396 | } |
| 397 | case LayerType::Floor : |
| 398 | { |
| 399 | auto floorQueueDescriptor = PolymorphicDowncast<const FloorQueueDescriptor*>(&descriptor); |
| 400 | return MakeWorkloadHelper<NeonFloorFloatWorkload, NullWorkload>(*floorQueueDescriptor, info); |
| 401 | } |
| 402 | case LayerType::FullyConnected : |
| 403 | { |
| 404 | auto fullyConnectedQueueDescriptor = PolymorphicDowncast<const FullyConnectedQueueDescriptor*>(&descriptor); |
| 405 | return std::make_unique<NeonFullyConnectedWorkload>(*fullyConnectedQueueDescriptor, |
| 406 | info, |
| 407 | m_MemoryManager->GetIntraLayerManager()); |
| 408 | } |
| 409 | case LayerType::Gather : |
| 410 | { |
| 411 | auto gatherQueueDescriptor = PolymorphicDowncast<const GatherQueueDescriptor*>(&descriptor); |
| 412 | return std::make_unique<NeonGatherWorkload>(*gatherQueueDescriptor, info); |
| 413 | } |
Teresa Charlin | bd22c7d | 2022-04-26 18:14:12 +0100 | [diff] [blame] | 414 | case LayerType::GatherNd : |
| 415 | { |
| 416 | auto gatherNdQueueDescriptor = PolymorphicDowncast<const GatherNdQueueDescriptor*>(&descriptor); |
| 417 | return std::make_unique<NeonGatherNdWorkload>(*gatherNdQueueDescriptor, info); |
| 418 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 419 | case LayerType::Input : |
| 420 | { |
| 421 | auto inputQueueDescriptor = PolymorphicDowncast<const InputQueueDescriptor*>(&descriptor); |
| 422 | return std::make_unique<CopyMemGenericWorkload>(*inputQueueDescriptor, info); |
| 423 | } |
| 424 | case LayerType::InstanceNormalization : |
| 425 | { |
| 426 | auto instanceNormalizationQueueDescriptor |
| 427 | = PolymorphicDowncast<const InstanceNormalizationQueueDescriptor*>(&descriptor); |
| 428 | return std::make_unique<NeonInstanceNormalizationWorkload>(*instanceNormalizationQueueDescriptor, info); |
| 429 | } |
| 430 | case LayerType::L2Normalization : |
| 431 | { |
| 432 | auto l2NormalizationQueueDescriptor |
| 433 | = PolymorphicDowncast<const L2NormalizationQueueDescriptor*>(&descriptor); |
| 434 | return MakeWorkloadHelper<NeonL2NormalizationFloatWorkload, NullWorkload> |
| 435 | (*l2NormalizationQueueDescriptor, info, m_MemoryManager->GetIntraLayerManager()); |
| 436 | } |
| 437 | case LayerType::LogSoftmax : |
| 438 | { |
| 439 | auto logSoftmaxQueueDescriptor = PolymorphicDowncast<const LogSoftmaxQueueDescriptor*>(&descriptor); |
| 440 | return std::make_unique<NeonLogSoftmaxWorkload>(*logSoftmaxQueueDescriptor, |
| 441 | info, |
| 442 | m_MemoryManager->GetIntraLayerManager()); |
| 443 | } |
| 444 | case LayerType::LogicalBinary : |
| 445 | { |
| 446 | auto logicalBinaryQueueDescriptor = PolymorphicDowncast<const LogicalBinaryQueueDescriptor*>(&descriptor); |
| 447 | |
| 448 | switch(logicalBinaryQueueDescriptor->m_Parameters.m_Operation) |
| 449 | { |
| 450 | case LogicalBinaryOperation::LogicalAnd: |
| 451 | return std::make_unique<NeonLogicalAndWorkload>(*logicalBinaryQueueDescriptor, info); |
| 452 | case LogicalBinaryOperation::LogicalOr: |
| 453 | return std::make_unique<NeonLogicalOrWorkload>(*logicalBinaryQueueDescriptor, info); |
| 454 | default: |
| 455 | return nullptr; |
| 456 | } |
| 457 | } |
| 458 | case LayerType::Lstm : |
| 459 | { |
| 460 | auto lstmQueueDescriptor = PolymorphicDowncast<const LstmQueueDescriptor*>(&descriptor); |
| 461 | return MakeWorkloadHelper<NeonLstmFloatWorkload, NullWorkload>(*lstmQueueDescriptor, info); |
| 462 | } |
| 463 | case LayerType::Maximum : |
| 464 | { |
| 465 | auto maximumQueueDescriptor = PolymorphicDowncast<const MaximumQueueDescriptor*>(&descriptor); |
| 466 | return std::make_unique<NeonMaximumWorkload>(*maximumQueueDescriptor, info); |
| 467 | } |
| 468 | case LayerType::Mean : |
| 469 | { |
| 470 | auto meanQueueDescriptor = PolymorphicDowncast<const MeanQueueDescriptor*>(&descriptor); |
| 471 | return std::make_unique<NeonMeanWorkload>(*meanQueueDescriptor, info); |
| 472 | } |
| 473 | case LayerType::MemCopy : |
| 474 | { |
| 475 | auto memCopyQueueDescriptor = PolymorphicDowncast<const MemCopyQueueDescriptor*>(&descriptor); |
| 476 | if (memCopyQueueDescriptor->m_Inputs.empty() || !memCopyQueueDescriptor->m_Inputs[0]) |
| 477 | { |
| 478 | throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemCopy workload"); |
| 479 | } |
| 480 | return MakeWorkloadHelper<CopyMemGenericWorkload, CopyMemGenericWorkload>(*memCopyQueueDescriptor, info); |
| 481 | } |
| 482 | case LayerType::MemImport : |
| 483 | { |
| 484 | auto memImportQueueDescriptor = PolymorphicDowncast<const MemImportQueueDescriptor*>(&descriptor); |
| 485 | if (memImportQueueDescriptor->m_Inputs.empty() || !memImportQueueDescriptor->m_Inputs[0]) |
| 486 | { |
| 487 | throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemImport workload"); |
| 488 | } |
| 489 | return std::make_unique<ImportMemGenericWorkload>(*memImportQueueDescriptor, info); |
| 490 | } |
| 491 | case LayerType::Minimum : |
| 492 | { |
| 493 | auto minimumQueueDescriptor = PolymorphicDowncast<const MinimumQueueDescriptor*>(&descriptor); |
| 494 | return std::make_unique<NeonMinimumWorkload>(*minimumQueueDescriptor, info); |
| 495 | } |
| 496 | case LayerType::Multiplication : |
| 497 | { |
| 498 | auto multiplicationQueueDescriptor = PolymorphicDowncast<const MultiplicationQueueDescriptor*>(&descriptor); |
| 499 | return std::make_unique<NeonMultiplicationWorkload>(*multiplicationQueueDescriptor, info); |
| 500 | } |
| 501 | case LayerType::Normalization : |
| 502 | { |
| 503 | auto normalizationQueueDescriptor = PolymorphicDowncast<const NormalizationQueueDescriptor*>(&descriptor); |
| 504 | return MakeWorkloadHelper<NeonNormalizationFloatWorkload, NullWorkload> |
| 505 | (*normalizationQueueDescriptor, info, m_MemoryManager->GetIntraLayerManager()); |
| 506 | } |
| 507 | case LayerType::Output : |
| 508 | { |
| 509 | auto outputQueueDescriptor = PolymorphicDowncast<const OutputQueueDescriptor*>(&descriptor); |
| 510 | return std::make_unique<CopyMemGenericWorkload>(*outputQueueDescriptor, info); |
| 511 | } |
| 512 | case LayerType::Pad : |
| 513 | { |
| 514 | auto padQueueDescriptor = PolymorphicDowncast<const PadQueueDescriptor*>(&descriptor); |
| 515 | return std::make_unique<NeonPadWorkload>(*padQueueDescriptor, info); |
| 516 | } |
| 517 | case LayerType::Permute : |
| 518 | { |
| 519 | auto permuteQueueDescriptor = PolymorphicDowncast<const PermuteQueueDescriptor*>(&descriptor); |
| 520 | return std::make_unique<NeonPermuteWorkload>(*permuteQueueDescriptor, info); |
| 521 | } |
| 522 | case LayerType::Pooling2d : |
| 523 | { |
| 524 | auto pooling2dQueueDescriptor = PolymorphicDowncast<const Pooling2dQueueDescriptor*>(&descriptor); |
| 525 | return std::make_unique<NeonPooling2dWorkload>(*pooling2dQueueDescriptor, info); |
| 526 | } |
Ryan OShea | 19e7942 | 2022-05-04 00:38:03 +0100 | [diff] [blame] | 527 | case LayerType::Pooling3d : |
| 528 | { |
| 529 | auto pooling3dQueueDescriptor = PolymorphicDowncast<const Pooling3dQueueDescriptor*>(&descriptor); |
| 530 | return std::make_unique<NeonPooling3dWorkload>(*pooling3dQueueDescriptor, info); |
| 531 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 532 | case LayerType::PreCompiled : |
| 533 | { |
| 534 | auto preCompiledQueueDescriptor = PolymorphicDowncast<const PreCompiledQueueDescriptor*>(&descriptor); |
| 535 | return MakeWorkloadHelper<NullWorkload, NullWorkload>(*preCompiledQueueDescriptor, info); |
| 536 | } |
| 537 | case LayerType::Prelu : |
| 538 | { |
| 539 | auto preluQueueDescriptor = PolymorphicDowncast<const PreluQueueDescriptor*>(&descriptor); |
| 540 | return std::make_unique<NeonPreluWorkload>(*preluQueueDescriptor, info); |
| 541 | } |
| 542 | case LayerType::QLstm : |
| 543 | { |
| 544 | auto qLstmQueueDescriptor = PolymorphicDowncast<const QLstmQueueDescriptor*>(&descriptor); |
| 545 | return std::make_unique<NeonQLstmWorkload>(*qLstmQueueDescriptor, info); |
| 546 | } |
| 547 | case LayerType::Quantize : |
| 548 | { |
| 549 | auto quantizeQueueDescriptor = PolymorphicDowncast<const QuantizeQueueDescriptor*>(&descriptor); |
| 550 | return std::make_unique<NeonQuantizeWorkload>(*quantizeQueueDescriptor, info); |
| 551 | } |
| 552 | case LayerType::QuantizedLstm : |
| 553 | { |
| 554 | auto quantizedLstmQueueDescriptor = PolymorphicDowncast<const QuantizedLstmQueueDescriptor*>(&descriptor); |
| 555 | return std::make_unique<NeonQuantizedLstmWorkload>(*quantizedLstmQueueDescriptor, info); |
| 556 | } |
| 557 | case LayerType::Rank : |
| 558 | { |
| 559 | auto rankQueueDescriptor = PolymorphicDowncast<const RankQueueDescriptor*>(&descriptor); |
| 560 | return std::make_unique<NeonRankWorkload>(*rankQueueDescriptor, info); |
| 561 | } |
| 562 | case LayerType::Reduce : |
| 563 | { |
| 564 | auto reduceQueueDescriptor = PolymorphicDowncast<const ReduceQueueDescriptor*>(&descriptor); |
| 565 | return std::make_unique<NeonReduceWorkload>(*reduceQueueDescriptor, info); |
| 566 | } |
| 567 | case LayerType::Reshape : |
| 568 | { |
| 569 | auto reshapeQueueDescriptor = PolymorphicDowncast<const ReshapeQueueDescriptor*>(&descriptor); |
| 570 | return std::make_unique<NeonReshapeWorkload>(*reshapeQueueDescriptor, info); |
| 571 | } |
| 572 | case LayerType::Resize : |
| 573 | { |
| 574 | auto resizeQueueDescriptor = PolymorphicDowncast<const ResizeQueueDescriptor*>(&descriptor); |
| 575 | return std::make_unique<NeonResizeWorkload>(*resizeQueueDescriptor, info); |
| 576 | } |
| 577 | case LayerType::Slice : |
| 578 | { |
| 579 | auto sliceQueueDescriptor = PolymorphicDowncast<const SliceQueueDescriptor*>(&descriptor); |
| 580 | return std::make_unique<NeonSliceWorkload>(*sliceQueueDescriptor, info); |
| 581 | } |
| 582 | case LayerType::Softmax : |
| 583 | { |
| 584 | auto softmaxQueueDescriptor = PolymorphicDowncast<const SoftmaxQueueDescriptor*>(&descriptor); |
| 585 | return std::make_unique<NeonSoftmaxWorkload>(*softmaxQueueDescriptor, |
| 586 | info, |
| 587 | m_MemoryManager->GetIntraLayerManager()); |
| 588 | } |
| 589 | case LayerType::SpaceToBatchNd : |
| 590 | { |
| 591 | auto spaceToBatchNdQueueDescriptor |
| 592 | = PolymorphicDowncast<const SpaceToBatchNdQueueDescriptor*>(&descriptor); |
| 593 | return std::make_unique<NeonSpaceToBatchNdWorkload>(*spaceToBatchNdQueueDescriptor, info); |
| 594 | } |
| 595 | case LayerType::SpaceToDepth : |
| 596 | { |
| 597 | auto spaceToDepthQueueDescriptor = PolymorphicDowncast<const SpaceToDepthQueueDescriptor*>(&descriptor); |
| 598 | return std::make_unique<NeonSpaceToDepthWorkload>(*spaceToDepthQueueDescriptor, info); |
| 599 | } |
| 600 | case LayerType::Splitter : |
| 601 | { |
| 602 | auto splitterQueueDescriptor = PolymorphicDowncast<const SplitterQueueDescriptor*>(&descriptor); |
| 603 | return std::make_unique<NeonSplitterWorkload>(*splitterQueueDescriptor, info); |
| 604 | } |
| 605 | case LayerType::Stack : |
| 606 | { |
| 607 | auto stackQueueDescriptor = PolymorphicDowncast<const StackQueueDescriptor*>(&descriptor); |
| 608 | return std::make_unique<NeonStackWorkload>(*stackQueueDescriptor, info); |
| 609 | } |
| 610 | case LayerType::StridedSlice : |
| 611 | { |
| 612 | auto stridedSliceQueueDescriptor = PolymorphicDowncast<const StridedSliceQueueDescriptor*>(&descriptor); |
| 613 | return std::make_unique<NeonStridedSliceWorkload>(*stridedSliceQueueDescriptor, info); |
| 614 | } |
| 615 | case LayerType::Subtraction : |
| 616 | { |
| 617 | auto subtractionQueueDescriptor = PolymorphicDowncast<const SubtractionQueueDescriptor*>(&descriptor); |
| 618 | return std::make_unique<NeonSubtractionWorkload>(*subtractionQueueDescriptor, info); |
| 619 | } |
David Monahan | 36e6eae | 2023-07-26 18:37:45 +0100 | [diff] [blame] | 620 | case LayerType::Tile: |
| 621 | { |
| 622 | auto tileQueueDescriptor = PolymorphicDowncast<const TileQueueDescriptor*>(&descriptor); |
| 623 | return std::make_unique<NeonTileWorkload>(*tileQueueDescriptor, info); |
| 624 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 625 | case LayerType::Transpose : |
| 626 | { |
| 627 | auto transposeQueueDescriptor = PolymorphicDowncast<const TransposeQueueDescriptor*>(&descriptor); |
| 628 | return std::make_unique<NeonTransposeWorkload>(*transposeQueueDescriptor, info); |
| 629 | } |
| 630 | case LayerType::TransposeConvolution2d : |
| 631 | { |
| 632 | auto transposeConvolution2dQueueDescriptor |
| 633 | = PolymorphicDowncast<const TransposeConvolution2dQueueDescriptor*>(&descriptor); |
| 634 | return std::make_unique<NeonTransposeConvolution2dWorkload>(*transposeConvolution2dQueueDescriptor, |
| 635 | info, |
| 636 | m_MemoryManager->GetIntraLayerManager()); |
| 637 | } |
Cathal Corbett | fd5bec4 | 2022-03-03 15:13:23 +0000 | [diff] [blame] | 638 | case LayerType::UnidirectionalSequenceLstm : |
| 639 | { |
| 640 | auto desc = PolymorphicDowncast<const UnidirectionalSequenceLstmQueueDescriptor*>(&descriptor); |
Mike Kelly | 1299496 | 2022-04-21 11:57:09 +0100 | [diff] [blame] | 641 | |
| 642 | if ((info.m_InputTensorInfos[0].GetDataType() == armnn::DataType::Float32) && |
| 643 | (info.m_InputTensorInfos[1].GetDataType() == armnn::DataType::Float32) && |
| 644 | (info.m_InputTensorInfos[2].GetDataType() == armnn::DataType::Float32) && |
| 645 | (info.m_OutputTensorInfos[0].GetDataType() == armnn::DataType::Float32) && |
| 646 | (info.m_OutputTensorInfos[1].GetDataType() == armnn::DataType::Float32) && |
| 647 | (info.m_OutputTensorInfos[2].GetDataType() == armnn::DataType::Float32)) |
| 648 | { |
| 649 | return std::make_unique<NeonUnidirectionalSequenceLstmFloatWorkload>(*desc, info); |
| 650 | } |
| 651 | else |
| 652 | { |
| 653 | return std::make_unique<NeonUnidirectionalSequenceLstmWorkload>(*desc, info); |
| 654 | } |
Cathal Corbett | fd5bec4 | 2022-03-03 15:13:23 +0000 | [diff] [blame] | 655 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 656 | default: |
| 657 | return nullptr; |
| 658 | } |
| 659 | } |
| 660 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 661 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateActivation(const ActivationQueueDescriptor& descriptor, |
| 662 | const WorkloadInfo& info) const |
| 663 | { |
Nattapat Chaimanowong | d4b7059 | 2018-10-12 11:21:49 +0100 | [diff] [blame] | 664 | return std::make_unique<NeonActivationWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 665 | } |
| 666 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 667 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateAddition(const AdditionQueueDescriptor& descriptor, |
| 668 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 669 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 670 | return std::make_unique<NeonAdditionWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 671 | } |
| 672 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 673 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateArgMinMax(const ArgMinMaxQueueDescriptor& descriptor, |
| 674 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 675 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 676 | return std::make_unique<NeonArgMinMaxWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 677 | } |
| 678 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 679 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateBatchNormalization( |
| 680 | const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 681 | { |
| 682 | return std::make_unique<NeonBatchNormalizationWorkload>(descriptor, info); |
| 683 | } |
| 684 | |
| 685 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateBatchToSpaceNd(const BatchToSpaceNdQueueDescriptor& descriptor, |
| 686 | const WorkloadInfo& info) const |
| 687 | { |
Mike Kelly | 5685802 | 2020-01-27 12:14:47 +0000 | [diff] [blame] | 688 | return std::make_unique<NeonBatchToSpaceNdWorkload>(descriptor, info); |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 689 | } |
| 690 | |
Sadik Armagan | 48f011e | 2021-04-21 10:50:34 +0100 | [diff] [blame] | 691 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateCast(const CastQueueDescriptor& descriptor, |
| 692 | const WorkloadInfo& info) const |
| 693 | { |
| 694 | return std::make_unique<NeonCastWorkload>(descriptor, info); |
| 695 | } |
| 696 | |
Teresa Charlin | e89dd69 | 2021-09-01 16:30:34 +0100 | [diff] [blame] | 697 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateChannelShuffle(const ChannelShuffleQueueDescriptor& descriptor, |
| 698 | const WorkloadInfo& info) const |
| 699 | { |
| 700 | return std::make_unique<NeonChannelShuffleWorkload>(descriptor, info); |
| 701 | } |
| 702 | |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 703 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateComparison(const ComparisonQueueDescriptor& descriptor, |
| 704 | const WorkloadInfo& info) const |
| 705 | { |
Teresa Charlin | cedd34f | 2020-03-30 11:17:30 +0100 | [diff] [blame] | 706 | return std::make_unique<NeonComparisonWorkload>(descriptor, info); |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 707 | } |
| 708 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 709 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConcat(const ConcatQueueDescriptor& descriptor, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 710 | const WorkloadInfo& info) const |
| 711 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 712 | return std::make_unique<NeonConcatWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 713 | } |
| 714 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 715 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConstant(const ConstantQueueDescriptor& descriptor, |
| 716 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 717 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 718 | return std::make_unique<NeonConstantWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 719 | } |
| 720 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 721 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp16ToFp32( |
| 722 | const ConvertFp16ToFp32QueueDescriptor& descriptor, |
| 723 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 724 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 725 | return std::make_unique<NeonConvertFp16ToFp32Workload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 726 | } |
| 727 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 728 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp32ToFp16( |
| 729 | const ConvertFp32ToFp16QueueDescriptor& descriptor, |
| 730 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 731 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 732 | return std::make_unique<NeonConvertFp32ToFp16Workload>(descriptor, info); |
Nikhil Raj | 9b46148 | 2019-07-03 15:58:31 +0100 | [diff] [blame] | 733 | } |
| 734 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 735 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConvolution2d( |
| 736 | const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 737 | { |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 738 | bool isFastMathEnabled = false; |
| 739 | if (m_ModelContextPtr) |
| 740 | { |
| 741 | if (m_ModelContextPtr.get() != nullptr) |
| 742 | { |
| 743 | auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get()); |
| 744 | if (modelOptions) |
| 745 | { |
| 746 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 747 | } |
| 748 | } |
| 749 | } |
| 750 | return std::make_unique<NeonConvolution2dWorkload>(descriptor, |
| 751 | info, |
| 752 | m_MemoryManager->GetIntraLayerManager(), |
| 753 | isFastMathEnabled); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 754 | } |
| 755 | |
Teresa Charlin | ec5f7d1 | 2021-10-22 17:15:00 +0100 | [diff] [blame] | 756 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConvolution3d( |
| 757 | const Convolution3dQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 758 | { |
| 759 | bool isFastMathEnabled = false; |
| 760 | if (m_ModelContextPtr) |
| 761 | { |
| 762 | if (m_ModelContextPtr.get() != nullptr) |
| 763 | { |
| 764 | auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get()); |
| 765 | if (modelOptions) |
| 766 | { |
| 767 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 768 | } |
| 769 | } |
| 770 | } |
| 771 | return std::make_unique<NeonConvolution3dWorkload>(descriptor, |
| 772 | info, |
| 773 | m_MemoryManager->GetIntraLayerManager(), |
| 774 | isFastMathEnabled); |
| 775 | } |
| 776 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 777 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDebug(const DebugQueueDescriptor& descriptor, |
| 778 | const WorkloadInfo& info) const |
| 779 | { |
| 780 | return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info); |
| 781 | } |
| 782 | |
Aron Virginas-Tar | 2f00b74 | 2019-09-30 13:28:08 +0100 | [diff] [blame] | 783 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor, |
| 784 | const WorkloadInfo& info) const |
| 785 | { |
| 786 | return std::make_unique<NeonDepthToSpaceWorkload>(descriptor, info); |
| 787 | } |
| 788 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 789 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthwiseConvolution2d( |
| 790 | const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 791 | { |
Nattapat Chaimanowong | 7714088 | 2018-10-17 11:12:19 +0100 | [diff] [blame] | 792 | return std::make_unique<NeonDepthwiseConvolutionWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 793 | } |
| 794 | |
Narumol Prangnawarat | 01961a7 | 2019-05-30 16:47:12 +0100 | [diff] [blame] | 795 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDequantize(const DequantizeQueueDescriptor& descriptor, |
| 796 | const WorkloadInfo& info) const |
| 797 | { |
| 798 | return std::make_unique<NeonDequantizeWorkload>(descriptor, info); |
| 799 | } |
| 800 | |
Narumol Prangnawarat | 94dd5d8 | 2019-01-23 18:06:26 +0000 | [diff] [blame] | 801 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDetectionPostProcess( |
| 802 | const armnn::DetectionPostProcessQueueDescriptor& descriptor, const armnn::WorkloadInfo& info) const |
| 803 | { |
James Conroy | d9fb6e2 | 2020-02-21 16:52:44 +0000 | [diff] [blame] | 804 | return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info); |
Narumol Prangnawarat | 94dd5d8 | 2019-01-23 18:06:26 +0000 | [diff] [blame] | 805 | } |
| 806 | |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 807 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateDivision( |
| 808 | const DivisionQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 809 | { |
Pablo Tello | e61f071 | 2020-01-23 10:37:17 +0000 | [diff] [blame] | 810 | return std::make_unique<NeonDivisionWorkload>(descriptor, info); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 811 | } |
| 812 | |
Sadik Armagan | ac47210 | 2020-03-24 09:54:36 +0000 | [diff] [blame] | 813 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateElementwiseUnary( |
| 814 | const ElementwiseUnaryQueueDescriptor& descriptor, const WorkloadInfo& info) const |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 815 | { |
Derek Lamberti | c77874a | 2020-04-28 13:34:56 +0100 | [diff] [blame] | 816 | switch(descriptor.m_Parameters.m_Operation) |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 817 | { |
Derek Lamberti | c77874a | 2020-04-28 13:34:56 +0100 | [diff] [blame] | 818 | case UnaryOperation::Abs: |
James Conroy | 177df1e | 2020-11-13 10:18:51 +0000 | [diff] [blame] | 819 | { |
| 820 | AbsQueueDescriptor absQueueDescriptor; |
| 821 | absQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 822 | absQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 823 | |
James Conroy | 177df1e | 2020-11-13 10:18:51 +0000 | [diff] [blame] | 824 | return std::make_unique<NeonAbsWorkload>(absQueueDescriptor, info); |
| 825 | } |
Teresa Charlin | 50de4fa | 2021-05-31 18:47:33 +0100 | [diff] [blame] | 826 | case UnaryOperation::Exp: |
| 827 | return std::make_unique<NeonExpWorkload>(descriptor, info); |
| 828 | case UnaryOperation::LogicalNot: |
| 829 | return std::make_unique<NeonLogicalNotWorkload>(descriptor, info); |
| 830 | case UnaryOperation::Log: |
| 831 | return std::make_unique<NeonLogWorkload>(descriptor, info); |
| 832 | case UnaryOperation::Neg: |
| 833 | return std::make_unique<NeonNegWorkload>(descriptor, info); |
Derek Lamberti | c77874a | 2020-04-28 13:34:56 +0100 | [diff] [blame] | 834 | case UnaryOperation::Rsqrt: |
James Conroy | 177df1e | 2020-11-13 10:18:51 +0000 | [diff] [blame] | 835 | { |
| 836 | RsqrtQueueDescriptor rsqrtQueueDescriptor; |
| 837 | rsqrtQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 838 | rsqrtQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 839 | |
James Conroy | 177df1e | 2020-11-13 10:18:51 +0000 | [diff] [blame] | 840 | return std::make_unique<NeonRsqrtWorkload>(rsqrtQueueDescriptor, info); |
| 841 | } |
Teresa Charlin | 50de4fa | 2021-05-31 18:47:33 +0100 | [diff] [blame] | 842 | case UnaryOperation::Sin: |
| 843 | return std::make_unique<NeonSinWorkload>(descriptor, info); |
Derek Lamberti | c77874a | 2020-04-28 13:34:56 +0100 | [diff] [blame] | 844 | default: |
| 845 | return nullptr; |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 846 | } |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 847 | } |
| 848 | |
Sadik Armagan | a792a05 | 2020-06-23 16:22:23 +0100 | [diff] [blame] | 849 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFill(const FillQueueDescriptor& descriptor, |
| 850 | const WorkloadInfo& info) const |
| 851 | { |
| 852 | return std::make_unique<NeonFillWorkload>(descriptor, info); |
| 853 | } |
| 854 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 855 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFloor(const FloorQueueDescriptor& descriptor, |
| 856 | const WorkloadInfo& info) const |
| 857 | { |
| 858 | return MakeWorkloadHelper<NeonFloorFloatWorkload, NullWorkload>(descriptor, info); |
| 859 | } |
| 860 | |
| 861 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateFullyConnected( |
| 862 | const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 863 | { |
Kevin May | be7e35c | 2020-04-29 17:05:05 +0100 | [diff] [blame] | 864 | return std::make_unique<NeonFullyConnectedWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager()); |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 865 | } |
| 866 | |
| 867 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateGather(const armnn::GatherQueueDescriptor& descriptor, |
| 868 | const armnn::WorkloadInfo& info) const |
| 869 | { |
Teresa Charlin | f540eb8 | 2020-04-10 19:24:55 +0100 | [diff] [blame] | 870 | return std::make_unique<NeonGatherWorkload>(descriptor, info); |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 871 | } |
| 872 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 873 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateInput(const InputQueueDescriptor& descriptor, |
| 874 | const WorkloadInfo& info) const |
| 875 | { |
| 876 | return std::make_unique<CopyMemGenericWorkload>(descriptor, info); |
| 877 | } |
| 878 | |
| 879 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateInstanceNormalization( |
| 880 | const InstanceNormalizationQueueDescriptor& descriptor, |
| 881 | const WorkloadInfo& info) const |
| 882 | { |
| 883 | return std::make_unique<NeonInstanceNormalizationWorkload>(descriptor, info); |
| 884 | } |
| 885 | |
| 886 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor, |
| 887 | const WorkloadInfo& info) const |
| 888 | { |
| 889 | return MakeWorkloadHelper<NeonL2NormalizationFloatWorkload, NullWorkload>(descriptor, info, |
| 890 | m_MemoryManager->GetIntraLayerManager()); |
| 891 | } |
| 892 | |
Keith Davis | 69e653f | 2020-07-02 11:49:26 +0100 | [diff] [blame] | 893 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLogSoftmax(const LogSoftmaxQueueDescriptor& descriptor, |
| 894 | const WorkloadInfo& info) const |
| 895 | { |
| 896 | return std::make_unique<NeonLogSoftmaxWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager()); |
| 897 | } |
| 898 | |
James Conroy | 177df1e | 2020-11-13 10:18:51 +0000 | [diff] [blame] | 899 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLogicalBinary(const LogicalBinaryQueueDescriptor& descriptor, |
| 900 | const WorkloadInfo& info) const |
| 901 | { |
| 902 | switch(descriptor.m_Parameters.m_Operation) |
| 903 | { |
| 904 | case LogicalBinaryOperation::LogicalAnd: |
| 905 | return std::make_unique<NeonLogicalAndWorkload>(descriptor, info); |
| 906 | case LogicalBinaryOperation::LogicalOr: |
| 907 | return std::make_unique<NeonLogicalOrWorkload>(descriptor, info); |
| 908 | default: |
| 909 | return nullptr; |
| 910 | } |
| 911 | } |
| 912 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 913 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLstm(const LstmQueueDescriptor& descriptor, |
| 914 | const WorkloadInfo& info) const |
| 915 | { |
| 916 | return MakeWorkloadHelper<NeonLstmFloatWorkload, NullWorkload>(descriptor, info); |
| 917 | } |
| 918 | |
| 919 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMaximum(const MaximumQueueDescriptor& descriptor, |
| 920 | const WorkloadInfo& info) const |
| 921 | { |
| 922 | return std::make_unique<NeonMaximumWorkload>(descriptor, info); |
| 923 | } |
| 924 | |
| 925 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor, |
| 926 | const WorkloadInfo& info) const |
| 927 | { |
| 928 | return std::make_unique<NeonMeanWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 929 | } |
| 930 | |
| 931 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMemCopy(const MemCopyQueueDescriptor& descriptor, |
| 932 | const WorkloadInfo& info) const |
| 933 | { |
| 934 | if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0]) |
| 935 | { |
| 936 | throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemCopy workload"); |
| 937 | } |
| 938 | |
Aron Virginas-Tar | a8e06ed | 2018-10-19 16:46:15 +0100 | [diff] [blame] | 939 | return MakeWorkloadHelper<CopyMemGenericWorkload, CopyMemGenericWorkload>(descriptor, info); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 940 | } |
| 941 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 942 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMemImport(const MemImportQueueDescriptor& descriptor, |
| 943 | const WorkloadInfo& info) const |
| 944 | { |
| 945 | if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0]) |
| 946 | { |
| 947 | throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemImport workload"); |
| 948 | } |
| 949 | |
| 950 | return std::make_unique<ImportMemGenericWorkload>(descriptor, info); |
| 951 | } |
| 952 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 953 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMinimum(const MinimumQueueDescriptor& descriptor, |
| 954 | const WorkloadInfo& info) const |
| 955 | { |
| 956 | return std::make_unique<NeonMinimumWorkload>(descriptor, info); |
| 957 | } |
| 958 | |
| 959 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMultiplication( |
| 960 | const MultiplicationQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 961 | { |
| 962 | return std::make_unique<NeonMultiplicationWorkload>(descriptor, info); |
| 963 | } |
| 964 | |
| 965 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateNormalization( |
| 966 | const NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 967 | { |
| 968 | return MakeWorkloadHelper<NeonNormalizationFloatWorkload, NullWorkload>(descriptor, info, |
| 969 | m_MemoryManager->GetIntraLayerManager()); |
| 970 | } |
| 971 | |
| 972 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateOutput(const OutputQueueDescriptor& descriptor, |
Keith Davis | 69e653f | 2020-07-02 11:49:26 +0100 | [diff] [blame] | 973 | const WorkloadInfo& info) const |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 974 | { |
| 975 | return std::make_unique<CopyMemGenericWorkload>(descriptor, info); |
| 976 | } |
| 977 | |
| 978 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor, |
| 979 | const WorkloadInfo& info) const |
| 980 | { |
| 981 | return std::make_unique<NeonPadWorkload>(descriptor, info); |
| 982 | } |
| 983 | |
| 984 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor, |
Keith Davis | 69e653f | 2020-07-02 11:49:26 +0100 | [diff] [blame] | 985 | const WorkloadInfo& info) const |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 986 | { |
| 987 | return std::make_unique<NeonPermuteWorkload>(descriptor, info); |
| 988 | } |
| 989 | |
| 990 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePooling2d(const Pooling2dQueueDescriptor& descriptor, |
Keith Davis | 69e653f | 2020-07-02 11:49:26 +0100 | [diff] [blame] | 991 | const WorkloadInfo& info) const |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 992 | { |
| 993 | return std::make_unique<NeonPooling2dWorkload>(descriptor, info); |
| 994 | } |
| 995 | |
| 996 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor, |
| 997 | const WorkloadInfo& info) const |
| 998 | { |
| 999 | return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info); |
| 1000 | } |
| 1001 | |
| 1002 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePrelu(const armnn::PreluQueueDescriptor &descriptor, |
| 1003 | const armnn::WorkloadInfo &info) const |
| 1004 | { |
| 1005 | return std::make_unique<NeonPreluWorkload>(descriptor, info); |
| 1006 | } |
| 1007 | |
James Conroy | cc34093 | 2020-05-12 18:08:52 +0100 | [diff] [blame] | 1008 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateQLstm(const QLstmQueueDescriptor& descriptor, |
| 1009 | const WorkloadInfo& info) const |
| 1010 | { |
| 1011 | return std::make_unique<NeonQLstmWorkload>(descriptor, info); |
| 1012 | } |
| 1013 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1014 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateQuantize(const QuantizeQueueDescriptor& descriptor, |
| 1015 | const WorkloadInfo& info) const |
| 1016 | { |
Kevin May | 9077473 | 2020-03-03 12:10:10 +0000 | [diff] [blame] | 1017 | return std::make_unique<NeonQuantizeWorkload>(descriptor, info); |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1018 | } |
| 1019 | |
| 1020 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& descriptor, |
| 1021 | const WorkloadInfo& info) const |
| 1022 | { |
| 1023 | return std::make_unique<NeonQuantizedLstmWorkload>(descriptor, info); |
| 1024 | } |
| 1025 | |
David Monahan | 97451b4 | 2020-12-03 09:48:06 +0000 | [diff] [blame] | 1026 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateRank(const RankQueueDescriptor& descriptor, |
| 1027 | const WorkloadInfo& info) const |
| 1028 | { |
| 1029 | return std::make_unique<NeonRankWorkload>(descriptor, info); |
| 1030 | } |
| 1031 | |
Sadik Armagan | a274748 | 2021-02-09 10:28:54 +0000 | [diff] [blame] | 1032 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateReduce(const ReduceQueueDescriptor& descriptor, |
| 1033 | const WorkloadInfo& info) const |
| 1034 | { |
| 1035 | return std::make_unique<NeonReduceWorkload>(descriptor, info); |
| 1036 | } |
| 1037 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1038 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor, |
| 1039 | const WorkloadInfo& info) const |
| 1040 | { |
| 1041 | return std::make_unique<NeonReshapeWorkload>(descriptor, info); |
| 1042 | } |
| 1043 | |
Aron Virginas-Tar | 169d2f1 | 2019-07-01 19:01:44 +0100 | [diff] [blame] | 1044 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateResize(const ResizeQueueDescriptor& descriptor, |
| 1045 | const WorkloadInfo& info) const |
| 1046 | { |
Ellen Norris-Thompson | 37e6868 | 2019-07-15 14:23:30 +0100 | [diff] [blame] | 1047 | return std::make_unique<NeonResizeWorkload>(descriptor, info); |
Aron Virginas-Tar | 169d2f1 | 2019-07-01 19:01:44 +0100 | [diff] [blame] | 1048 | } |
| 1049 | |
josh minor | 036f02d | 2019-11-15 14:53:22 -0600 | [diff] [blame] | 1050 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSlice(const SliceQueueDescriptor& descriptor, |
| 1051 | const WorkloadInfo& info) const |
| 1052 | { |
| 1053 | return std::make_unique<NeonSliceWorkload>(descriptor, info); |
| 1054 | } |
| 1055 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1056 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor, |
| 1057 | const WorkloadInfo& info) const |
Sadik Armagan | 581742d | 2019-08-12 14:11:37 +0100 | [diff] [blame] | 1058 | { |
Sadik Armagan | be88a57 | 2020-04-30 11:39:37 +0100 | [diff] [blame] | 1059 | return std::make_unique<NeonSoftmaxWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager()); |
Sadik Armagan | 581742d | 2019-08-12 14:11:37 +0100 | [diff] [blame] | 1060 | } |
| 1061 | |
Mike Kelly | 0be3a88 | 2020-01-24 11:27:50 +0000 | [diff] [blame] | 1062 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor, |
| 1063 | const WorkloadInfo& info) const |
| 1064 | { |
| 1065 | return std::make_unique<NeonSpaceToBatchNdWorkload>(descriptor, info); |
| 1066 | } |
| 1067 | |
| 1068 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSpaceToDepth(const SpaceToDepthQueueDescriptor& descriptor, |
| 1069 | const WorkloadInfo& info) const |
narpra01 | b89b05f | 2019-01-16 09:53:09 +0000 | [diff] [blame] | 1070 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1071 | return std::make_unique<NeonSpaceToDepthWorkload>(descriptor, info); |
| 1072 | } |
| 1073 | |
| 1074 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSplitter(const SplitterQueueDescriptor& descriptor, |
| 1075 | const WorkloadInfo& info) const |
| 1076 | { |
| 1077 | return std::make_unique<NeonSplitterWorkload>(descriptor, info); |
narpra01 | b89b05f | 2019-01-16 09:53:09 +0000 | [diff] [blame] | 1078 | } |
| 1079 | |
Matthew Jackson | 87f65ea | 2019-08-01 10:01:34 +0100 | [diff] [blame] | 1080 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateStack(const StackQueueDescriptor& descriptor, |
| 1081 | const WorkloadInfo& info) const |
| 1082 | { |
| 1083 | return std::make_unique<NeonStackWorkload>(descriptor, info); |
| 1084 | } |
| 1085 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1086 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor, |
| 1087 | const WorkloadInfo& info) const |
James Conroy | d47a064 | 2019-09-17 14:22:06 +0100 | [diff] [blame] | 1088 | { |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1089 | return std::make_unique<NeonStridedSliceWorkload>(descriptor, info); |
| 1090 | } |
| 1091 | |
| 1092 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateSubtraction( |
| 1093 | const SubtractionQueueDescriptor& descriptor, const WorkloadInfo& info) const |
| 1094 | { |
| 1095 | return std::make_unique<NeonSubtractionWorkload>(descriptor, info); |
| 1096 | } |
| 1097 | |
Mike Kelly | c9ea45a | 2020-02-28 18:11:58 +0000 | [diff] [blame] | 1098 | std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateTranspose(const TransposeQueueDescriptor& descriptor, |
| 1099 | const WorkloadInfo& info) const |
| 1100 | { |
| 1101 | return std::make_unique<NeonTransposeWorkload>(descriptor, info); |
| 1102 | } |
| 1103 | |
Sadik Armagan | 0d4863d | 2019-10-09 14:26:32 +0100 | [diff] [blame] | 1104 | std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateTransposeConvolution2d( |
| 1105 | const TransposeConvolution2dQueueDescriptor &descriptor, |
| 1106 | const WorkloadInfo &info) const |
| 1107 | { |
| 1108 | return std::make_unique<NeonTransposeConvolution2dWorkload>(descriptor, info, |
| 1109 | m_MemoryManager->GetIntraLayerManager()); |
James Conroy | d47a064 | 2019-09-17 14:22:06 +0100 | [diff] [blame] | 1110 | } |
| 1111 | |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 1112 | } // namespace armnn |