telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. 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 | // |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 5 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 6 | #include "Network.hpp" |
| 7 | #include "Graph.hpp" |
| 8 | #include "Layer.hpp" |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 9 | #include "DeviceSpec.hpp" |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 10 | #include "Optimizer.hpp" |
Derek Lamberti | ff05cc5 | 2019-04-26 13:05:17 +0100 | [diff] [blame] | 11 | #include "SubgraphViewSelector.hpp" |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 12 | #include "BackendSettings.hpp" |
David Beck | ac42efd | 2018-09-26 17:41:13 +0100 | [diff] [blame] | 13 | #include "optimizations/All.hpp" |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 14 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 15 | #include <backendsCommon/CpuTensorHandle.hpp> |
| 16 | #include <backendsCommon/WorkloadFactory.hpp> |
Matteo Martincigh | e5b8eb9 | 2019-11-28 15:45:42 +0000 | [diff] [blame] | 17 | #include <armnn/backends/IBackendInternal.hpp> |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 18 | #include <backendsCommon/TensorHandleFactoryRegistry.hpp> |
David Beck | ac42efd | 2018-09-26 17:41:13 +0100 | [diff] [blame] | 19 | |
| 20 | #include <armnn/Exceptions.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 21 | #include <armnn/Utils.hpp> |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 22 | #include <armnn/TypesUtils.hpp> |
Matteo Martincigh | c601aa6 | 2019-10-29 15:03:22 +0000 | [diff] [blame] | 23 | #include <armnn/BackendRegistry.hpp> |
Matthew Bentham | f48afc6 | 2020-01-15 17:55:08 +0000 | [diff] [blame] | 24 | #include <armnn/Logging.hpp> |
Jan Eilers | 8eb2560 | 2020-03-09 12:13:48 +0000 | [diff] [blame] | 25 | #include <armnn/utility/IgnoreUnused.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 26 | |
Jan Eilers | 99d9d4a | 2019-11-06 10:02:16 +0000 | [diff] [blame] | 27 | #include <ProfilingService.hpp> |
| 28 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 29 | #include <fcntl.h> |
| 30 | #include <algorithm> |
| 31 | #include <fstream> |
| 32 | #include <memory> |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 33 | #include <vector> |
| 34 | #include <algorithm> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 35 | |
| 36 | #include <boost/assert.hpp> |
| 37 | #include <boost/format.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 38 | #include <boost/numeric/conversion/converter_policies.hpp> |
| 39 | #include <boost/cast.hpp> |
| 40 | |
| 41 | namespace armnn |
| 42 | { |
| 43 | |
| 44 | armnn::INetwork* INetwork::CreateRaw() |
| 45 | { |
| 46 | return new Network(); |
| 47 | } |
| 48 | |
| 49 | armnn::INetworkPtr INetwork::Create() |
| 50 | { |
| 51 | return INetworkPtr(CreateRaw(), &INetwork::Destroy); |
| 52 | } |
| 53 | |
| 54 | void INetwork::Destroy(INetwork* network) |
| 55 | { |
| 56 | delete boost::polymorphic_downcast<Network*>(network); |
| 57 | } |
| 58 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 59 | void IOptimizedNetwork::Destroy(IOptimizedNetwork* network) |
| 60 | { |
| 61 | delete boost::polymorphic_downcast<OptimizedNetwork*>(network); |
| 62 | } |
| 63 | |
| 64 | Status OptimizedNetwork::PrintGraph() |
| 65 | { |
| 66 | m_Graph->Print(); |
| 67 | return Status::Success; |
| 68 | } |
| 69 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 70 | Status OptimizedNetwork::SerializeToDot(std::ostream& stream) const |
| 71 | { |
| 72 | return m_Graph->SerializeToDot(stream); |
| 73 | } |
| 74 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 75 | void ReportError(const std::string& errorMessage, |
| 76 | Optional<std::vector<std::string>&> errorMessages) |
| 77 | { |
| 78 | std::stringstream fullErrorMessage; |
| 79 | fullErrorMessage << "ERROR: " << errorMessage; |
Derek Lamberti | 0844697 | 2019-11-26 16:38:31 +0000 | [diff] [blame] | 80 | ARMNN_LOG(warning) << fullErrorMessage.str(); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 81 | if (errorMessages) |
| 82 | { |
| 83 | errorMessages.value().push_back(fullErrorMessage.str()); |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | void ReportWarning(const std::string& warningMessage, |
| 88 | Optional<std::vector<std::string>&> warningMessages) |
| 89 | { |
| 90 | std::stringstream fullWarningMessage; |
| 91 | fullWarningMessage << "WARNING: " << warningMessage; |
Derek Lamberti | 0844697 | 2019-11-26 16:38:31 +0000 | [diff] [blame] | 92 | ARMNN_LOG(warning) << fullWarningMessage.str(); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 93 | if (warningMessages) |
| 94 | { |
| 95 | warningMessages.value().push_back(fullWarningMessage.str()); |
| 96 | } |
| 97 | } |
| 98 | |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 99 | OptimizationResult ReturnWithError(OptimizationResult res, |
| 100 | const Layer* layer, |
| 101 | const BackendSettings& backendSettings, |
| 102 | Optional<std::vector<std::string>&> errMessages) |
| 103 | { |
| 104 | std::stringstream failureMsg; |
| 105 | failureMsg << "Layer of type " << GetLayerTypeAsCString(layer->GetType()) |
| 106 | << " is not supported on any preferred backend " << backendSettings.m_PreferredBackends; |
| 107 | ReportError(failureMsg.str(), errMessages); |
| 108 | |
| 109 | res.m_Error = true; |
| 110 | return res; |
| 111 | } |
| 112 | |
| 113 | |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 114 | bool CheckScaleSetOnQuantizedType(Layer* layer, Optional<std::vector<std::string>&> errMessages) |
| 115 | { |
| 116 | bool noErrors = true; |
| 117 | unsigned int numOutputs = layer->GetNumOutputSlots(); |
| 118 | for (unsigned int i = 0; i < numOutputs; i++) { |
David Monahan | b855470 | 2019-04-25 16:03:38 +0100 | [diff] [blame] | 119 | OutputSlot& outputSlot = layer->GetOutputSlot(i); |
| 120 | TensorInfo info = outputSlot.GetTensorInfo(); |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 121 | if (DataType::QAsymmU8 == info.GetDataType()) { |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 122 | if (0.f == info.GetQuantizationScale()) { |
| 123 | noErrors = false; |
| 124 | std::stringstream ss; |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 125 | ss << "output " << i << " of layer " << GetLayerTypeAsCString(layer->GetType()) |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 126 | << " (" << layer->GetNameStr() << ") is of type" |
| 127 | << " Quantized 8 bit but its scale parameter has not been set"; |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 128 | ReportError(ss.str(), errMessages); |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 129 | } |
David Monahan | b855470 | 2019-04-25 16:03:38 +0100 | [diff] [blame] | 130 | // Softmax under QuantisedAsymm8 must always be scale (1.0f/256.0f) and offset 0 |
| 131 | if ((info.GetQuantizationScale() != (1.0f / 256.0f) || |
| 132 | info.GetQuantizationOffset() != 0) && |
| 133 | layer->GetType() == armnn::LayerType::Softmax) |
| 134 | { |
| 135 | std::stringstream ss; |
| 136 | ss << "Quantization parameters for Softmax layer (Scale: " << |
| 137 | info.GetQuantizationScale() << " and Offset: " << info.GetQuantizationOffset() << |
| 138 | ") are incorrect and have been updated to Scale: 0.00390625 and Offset: 0"; |
Derek Lamberti | 0844697 | 2019-11-26 16:38:31 +0000 | [diff] [blame] | 139 | ARMNN_LOG(warning) << ss.str(); |
David Monahan | b855470 | 2019-04-25 16:03:38 +0100 | [diff] [blame] | 140 | info.SetQuantizationScale((1.0f /256.0f)); |
| 141 | info.SetQuantizationOffset(0); |
| 142 | outputSlot.SetTensorInfo(info); |
| 143 | } |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 144 | } |
| 145 | } |
| 146 | return noErrors; |
| 147 | } |
| 148 | |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 149 | OptimizationResult AttemptBackendAssignment(BackendSettings& backendSettings, |
| 150 | Graph& graph, |
| 151 | Layer* layer, |
| 152 | BackendId backend, |
| 153 | DataType dataTypeIn, |
| 154 | DataType dataTypeOut, |
| 155 | const std::vector<BackendId>& availablePreferredBackends, |
| 156 | std::string& reasonIfUnsupported, |
| 157 | Optional<std::vector<std::string>&> errMessages) |
| 158 | { |
| 159 | OptimizationResult result; |
| 160 | |
| 161 | // Helper lambda to compose meaningful error message before returning with error |
| 162 | auto ReturnError = [&](const Layer* layer) |
| 163 | { |
| 164 | return ReturnWithError(result, layer, backendSettings, errMessages); |
| 165 | }; |
| 166 | |
| 167 | // need to set the compute device on the layer |
| 168 | // before we can check if it is supported |
| 169 | layer->SetBackendId(backend); |
| 170 | if (!IWorkloadFactory::IsLayerSupported(*layer, EmptyOptional(), reasonIfUnsupported)) |
| 171 | { |
| 172 | if (dataTypeIn == DataType::Float16 || dataTypeOut == DataType::Float16) |
| 173 | { |
| 174 | if (IWorkloadFactory::IsLayerSupported(*layer, DataType::Float32, reasonIfUnsupported) |
| 175 | && layer->GetType() != LayerType::ConvertFp32ToFp16 |
| 176 | && layer->GetType() != LayerType::ConvertFp16ToFp32) |
| 177 | { |
| 178 | // Insert FP16 -> FP32 conversion layer before current layer |
| 179 | std::vector<ConvertFp16ToFp32Layer*> convertFp16ToFp32Layers; |
| 180 | if (dataTypeIn == DataType::Float16) |
| 181 | { |
| 182 | convertFp16ToFp32Layers = |
| 183 | InsertConvertFp16ToFp32LayersBefore(graph, *layer); |
| 184 | } |
| 185 | |
| 186 | // Insert FP32 -> FP16 conversion layer after current layer |
| 187 | std::vector<ConvertFp32ToFp16Layer*> convertFp32ToFp16Layers; |
| 188 | if (dataTypeOut == DataType::Float16) |
| 189 | { |
| 190 | convertFp32ToFp16Layers = |
| 191 | InsertConvertFp32ToFp16LayersAfter(graph, *layer); |
| 192 | } |
| 193 | |
| 194 | // Assign a supported backend to the newly introduced conversion layers |
| 195 | auto AssignFirstSupportedBackend = [&](Layer* layer, BackendId preferredBackend) |
| 196 | { |
| 197 | bool supportedBackendFound = false; |
| 198 | std::string reasonIfUnsupported; |
| 199 | |
| 200 | // Try preferred backend first |
| 201 | layer->SetBackendId(preferredBackend); |
| 202 | if (IWorkloadFactory::IsLayerSupported(*layer, |
| 203 | EmptyOptional(), |
| 204 | reasonIfUnsupported)) |
| 205 | { |
| 206 | supportedBackendFound = true; |
| 207 | } |
| 208 | else |
| 209 | { |
| 210 | for (const auto& backend : availablePreferredBackends) |
| 211 | { |
| 212 | // Skip preferred backend (we already determined that it is not supported) |
| 213 | if (backend == preferredBackend) |
| 214 | { |
| 215 | continue; |
| 216 | } |
| 217 | |
| 218 | layer->SetBackendId(backend); |
| 219 | if (IWorkloadFactory::IsLayerSupported(*layer, |
| 220 | EmptyOptional(), |
| 221 | reasonIfUnsupported)) |
| 222 | { |
| 223 | supportedBackendFound = true; |
| 224 | break; |
| 225 | } |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | return supportedBackendFound; |
| 230 | }; |
| 231 | |
| 232 | for (ConvertFp16ToFp32Layer* convertLayer : convertFp16ToFp32Layers) |
| 233 | { |
| 234 | if (!AssignFirstSupportedBackend(convertLayer, backend)) |
| 235 | { |
| 236 | return ReturnError(convertLayer); |
| 237 | } |
| 238 | } |
| 239 | |
| 240 | for (ConvertFp32ToFp16Layer* convertLayer : convertFp32ToFp16Layers) |
| 241 | { |
| 242 | if (!AssignFirstSupportedBackend(convertLayer, backend)) |
| 243 | { |
| 244 | return ReturnError(convertLayer); |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | return result; |
| 249 | } |
| 250 | } |
Narumol Prangnawarat | bc7ffb5 | 2020-03-20 15:01:01 +0000 | [diff] [blame^] | 251 | else if (dataTypeIn == DataType::BFloat16 || dataTypeOut == DataType::BFloat16) |
| 252 | { |
| 253 | if (IWorkloadFactory::IsLayerSupported(*layer, DataType::Float32, reasonIfUnsupported) |
| 254 | && layer->GetType() != LayerType::ConvertFp32ToBf16 |
| 255 | && layer->GetType() != LayerType::ConvertBf16ToFp32) |
| 256 | { |
| 257 | // Insert BF16 -> FP32 conversion layer before current layer |
| 258 | std::vector<ConvertBf16ToFp32Layer*> convertBf16ToFp32Layers; |
| 259 | if (dataTypeIn == DataType::BFloat16) |
| 260 | { |
| 261 | convertBf16ToFp32Layers = |
| 262 | InsertConvertBf16ToFp32LayersBefore(graph, *layer); |
| 263 | } |
| 264 | |
| 265 | // Insert FP32 -> BF16 conversion layer after current layer |
| 266 | std::vector<ConvertFp32ToBf16Layer*> convertFp32ToBf16Layers; |
| 267 | if (dataTypeOut == DataType::BFloat16) |
| 268 | { |
| 269 | convertFp32ToBf16Layers = |
| 270 | InsertConvertFp32ToBf16LayersAfter(graph, *layer); |
| 271 | } |
| 272 | |
| 273 | // Assign a supported backend to the newly introduced conversion layers |
| 274 | auto AssignFirstSupportedBackend = [&](Layer* layer, BackendId preferredBackend) |
| 275 | { |
| 276 | bool supportedBackendFound = false; |
| 277 | std::string reasonIfUnsupported; |
| 278 | |
| 279 | // Try preferred backend first |
| 280 | layer->SetBackendId(preferredBackend); |
| 281 | if (IWorkloadFactory::IsLayerSupported(*layer, |
| 282 | EmptyOptional(), |
| 283 | reasonIfUnsupported)) |
| 284 | { |
| 285 | supportedBackendFound = true; |
| 286 | } |
| 287 | else |
| 288 | { |
| 289 | for (const auto& backend : availablePreferredBackends) |
| 290 | { |
| 291 | // Skip preferred backend (we already determined that it is not supported) |
| 292 | if (backend == preferredBackend) |
| 293 | { |
| 294 | continue; |
| 295 | } |
| 296 | |
| 297 | layer->SetBackendId(backend); |
| 298 | if (IWorkloadFactory::IsLayerSupported(*layer, |
| 299 | EmptyOptional(), |
| 300 | reasonIfUnsupported)) |
| 301 | { |
| 302 | supportedBackendFound = true; |
| 303 | break; |
| 304 | } |
| 305 | } |
| 306 | } |
| 307 | |
| 308 | return supportedBackendFound; |
| 309 | }; |
| 310 | |
| 311 | for (ConvertBf16ToFp32Layer* convertLayer : convertBf16ToFp32Layers) |
| 312 | { |
| 313 | if (!AssignFirstSupportedBackend(convertLayer, backend)) |
| 314 | { |
| 315 | return ReturnError(convertLayer); |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | for (ConvertFp32ToBf16Layer* convertLayer : convertFp32ToBf16Layers) |
| 320 | { |
| 321 | if (!AssignFirstSupportedBackend(convertLayer, backend)) |
| 322 | { |
| 323 | return ReturnError(convertLayer); |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | return result; |
| 328 | } |
| 329 | } |
| 330 | |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 331 | std::stringstream warningMsg; |
| 332 | warningMsg << "Layer of type " << GetLayerTypeAsCString(layer->GetType()) |
| 333 | << " is not supported on requested backend " << layer->GetBackendId().Get() |
| 334 | << " for input data type " << GetDataTypeName(dataTypeIn) |
| 335 | << " and output data type " << GetDataTypeName(dataTypeOut) |
| 336 | << " (reason: " << reasonIfUnsupported |
| 337 | << "), falling back to the next backend."; |
| 338 | ReportWarning(warningMsg.str(), errMessages); |
| 339 | |
| 340 | return OptimizationResult(true, false); |
| 341 | } |
| 342 | else |
| 343 | { |
| 344 | return result; |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 349 | OptimizationResult AssignBackends(OptimizedNetwork* optNetObjPtr, |
| 350 | BackendSettings& backendSettings, |
| 351 | Graph::Iterator& firstLayer, |
| 352 | Graph::Iterator& lastLayer, |
| 353 | Optional<std::vector<std::string>&> errMessages) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 354 | { |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 355 | OptimizationResult result; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 356 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 357 | // Helper lambda to compose meaningful error message before returning with error |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 358 | auto ReturnError = [&](const Layer* layer) |
| 359 | { |
| 360 | return ReturnWithError(result, layer, backendSettings, errMessages); |
| 361 | }; |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 362 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 363 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 364 | auto availablePreferredBackends = backendSettings.GetAvailablePreferredBackends(); |
| 365 | if (availablePreferredBackends.empty()) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 366 | { |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 367 | std::stringstream failureMsg; |
| 368 | failureMsg << "No preferred backends are available"; |
| 369 | ReportError(failureMsg.str(), errMessages); |
| 370 | |
| 371 | result.m_Error = true; |
| 372 | return result; |
| 373 | } |
| 374 | |
| 375 | for (auto it = firstLayer; it != lastLayer; ++it) |
| 376 | { |
| 377 | auto layer = *it; |
Aron Virginas-Tar | 87972be | 2019-11-13 15:16:28 +0000 | [diff] [blame] | 378 | |
| 379 | DataType dataTypeIn = layer->GetNumInputSlots() == 0 ? DataType::Float32 : |
| 380 | layer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo().GetDataType(); |
| 381 | DataType dataTypeOut = layer->GetNumOutputSlots() == 0 ? DataType::Float32 : |
| 382 | layer->GetOutputSlot(0).GetTensorInfo().GetDataType(); |
| 383 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 384 | std::string reasonIfUnsupported; |
| 385 | bool found = false; |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 386 | if (!CheckScaleSetOnQuantizedType(layer, errMessages)) |
| 387 | { |
| 388 | // don't bomb immediately, find all the quantized outputs |
| 389 | // which haven't had a scale set and report them all back. |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 390 | result.m_Error = true; |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 391 | } |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 392 | |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 393 | // First try assign layer to hint backend |
| 394 | if (layer->GetBackendHint().has_value() && |
| 395 | backendSettings.IsBackendSupported(layer->GetBackendHint().value()) && |
| 396 | AttemptBackendAssignment(backendSettings, |
| 397 | optNetObjPtr->GetGraph(), |
| 398 | layer, |
| 399 | layer->GetBackendHint().value(), |
| 400 | dataTypeIn, |
| 401 | dataTypeOut, |
| 402 | availablePreferredBackends, |
| 403 | reasonIfUnsupported, |
| 404 | errMessages).IsOk()) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 405 | { |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 406 | found = true; |
| 407 | backendSettings.m_SelectedBackends.insert(layer->GetBackendHint().value()); |
| 408 | } |
| 409 | else |
| 410 | { |
| 411 | // Try assign layer to prefered list of backends |
| 412 | for (const auto& backend : availablePreferredBackends) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 413 | { |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 414 | if (layer->GetBackendHint().has_value() && |
| 415 | layer->GetBackendHint().value() == backend) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 416 | { |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 417 | continue; //Don't re-test the backend hint |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 418 | } |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 419 | |
| 420 | OptimizationResult res = AttemptBackendAssignment(backendSettings, |
| 421 | optNetObjPtr->GetGraph(), |
| 422 | layer, |
| 423 | backend, |
| 424 | dataTypeIn, |
| 425 | dataTypeOut, |
| 426 | availablePreferredBackends, |
| 427 | reasonIfUnsupported, |
| 428 | errMessages); |
| 429 | |
| 430 | if (res.IsOk()) |
| 431 | { |
| 432 | found = true; |
| 433 | backendSettings.m_SelectedBackends.insert(backend); |
| 434 | break; |
| 435 | } |
| 436 | else if (res.IsError()) |
| 437 | { |
| 438 | return res; // Cannot continue. |
| 439 | // Note: we don't need to log the error as it would already |
| 440 | // be logged in AttemptBackendAssignment(). |
| 441 | } |
| 442 | else |
| 443 | { |
| 444 | BOOST_ASSERT_MSG(res.IsWarningOnly(), "OptimizationResult in unexpected state."); |
| 445 | } |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 446 | } |
| 447 | } |
| 448 | |
| 449 | // If the layer is unsupported by any devices, log and return a null network. |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 450 | if (!found) |
| 451 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 452 | // NOTE: if the layer is not an operation queue type AND we have not got CpuRef as a |
| 453 | // fallback we should set the compute device on the layer to CpuRef (these are not |
| 454 | // available as accelerated operations, or are only available under certain |
| 455 | // conditions, currently they comprise MemCopy, Constant, Permute) |
| 456 | armnn::LayerType layerType = layer->GetType(); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 457 | if (!backendSettings.IsCpuRefUsed() && (layerType == armnn::LayerType::MemCopy || |
| 458 | layerType == armnn::LayerType::Constant || |
| 459 | layerType == armnn::LayerType::Permute)) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 460 | { |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 461 | BackendId cpuBackendId(armnn::Compute::CpuRef); |
| 462 | layer->SetBackendId(cpuBackendId); |
| 463 | backendSettings.m_SelectedBackends.insert(cpuBackendId); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 464 | } |
| 465 | else |
| 466 | { |
Derek Lamberti | 4a9e24b | 2020-01-03 16:53:38 +0000 | [diff] [blame] | 467 | return ReturnError(layer); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 468 | } |
| 469 | } |
| 470 | } |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 471 | |
| 472 | return result; |
| 473 | } |
| 474 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 475 | OptimizationResult AssignBackends(OptimizedNetwork* optNetObjPtr, |
| 476 | BackendSettings& backendSettings, |
Derek Lamberti | ff05cc5 | 2019-04-26 13:05:17 +0100 | [diff] [blame] | 477 | SubgraphView& subgraph, |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 478 | Optional<std::vector<std::string>&> errMessages) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 479 | { |
Derek Lamberti | ff05cc5 | 2019-04-26 13:05:17 +0100 | [diff] [blame] | 480 | Graph::Iterator firstLayer = subgraph.begin(); |
| 481 | Graph::Iterator lastLayer = subgraph.end(); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 482 | return AssignBackends(optNetObjPtr, |
| 483 | backendSettings, |
| 484 | firstLayer, |
| 485 | lastLayer, |
| 486 | errMessages); |
| 487 | } |
| 488 | |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 489 | BackendsMap CreateSupportedBackends(TensorHandleFactoryRegistry& handleFactoryRegistry, |
| 490 | BackendSettings& backendSettings) |
| 491 | { |
| 492 | BackendsMap backends; |
| 493 | auto const& backendRegistry = BackendRegistryInstance(); |
| 494 | for (auto&& selectedBackend : backendSettings.m_SupportedBackends) |
| 495 | { |
| 496 | auto backendFactory = backendRegistry.GetFactory(selectedBackend); |
| 497 | auto backendObjPtr = backendFactory(); |
| 498 | BOOST_ASSERT(backendObjPtr); |
| 499 | |
| 500 | backendObjPtr->RegisterTensorHandleFactories(handleFactoryRegistry); |
| 501 | |
| 502 | backends[backendObjPtr->GetId()] = std::move(backendObjPtr); |
| 503 | } |
| 504 | |
| 505 | return backends; |
| 506 | } |
| 507 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 508 | OptimizationResult ApplyBackendOptimizations(OptimizedNetwork* optNetObjPtr, |
| 509 | BackendSettings& backendSettings, |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 510 | BackendsMap& backends, |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 511 | Optional<std::vector<std::string>&> errMessages) |
| 512 | { |
| 513 | BOOST_ASSERT(optNetObjPtr); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 514 | |
| 515 | OptimizationResult result; |
| 516 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 517 | // Get the optimized graph |
| 518 | Graph& optGraph = optNetObjPtr->GetGraph(); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 519 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 520 | // Run backend specific optimizations |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 521 | for (auto&& selectedBackend : backendSettings.m_SelectedBackends) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 522 | { |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 523 | auto backendObjPtr = backends.find(selectedBackend)->second.get(); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 524 | BOOST_ASSERT(backendObjPtr); |
| 525 | |
| 526 | // Select sub-graphs based on backend |
Derek Lamberti | ff05cc5 | 2019-04-26 13:05:17 +0100 | [diff] [blame] | 527 | SubgraphViewSelector::Subgraphs subgraphs = |
Rob Hughes | 65c3226 | 2019-07-23 15:33:39 +0100 | [diff] [blame] | 528 | SubgraphViewSelector::SelectSubgraphs(optGraph, |
Matteo Martincigh | 602af09 | 2019-05-01 10:31:27 +0100 | [diff] [blame] | 529 | // Select layers assigned to the requested backend |
| 530 | [&backendObjPtr](const Layer& layer) |
| 531 | { |
| 532 | return layer.GetType() != LayerType::Input && |
| 533 | layer.GetType() != LayerType::Output && |
| 534 | layer.GetBackendId() == backendObjPtr->GetId(); |
| 535 | }); |
Derek Lamberti | ff05cc5 | 2019-04-26 13:05:17 +0100 | [diff] [blame] | 536 | if (subgraphs.empty()) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 537 | { |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 538 | // No sub-graphs found, try with next selected backend |
| 539 | continue; |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 540 | } |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 541 | |
| 542 | // Try to optimize each sub-graph |
Derek Lamberti | ff05cc5 | 2019-04-26 13:05:17 +0100 | [diff] [blame] | 543 | for (auto& subgraph : subgraphs) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 544 | { |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 545 | // Try to optimize the current sub-graph |
Matteo Martincigh | 8492433 | 2019-05-09 12:46:16 +0100 | [diff] [blame] | 546 | OptimizationViews optimizationViews = backendObjPtr->OptimizeSubgraphView(*subgraph); |
| 547 | BOOST_ASSERT(optimizationViews.Validate(*subgraph)); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 548 | |
| 549 | // Optimization attempted, check the resulting optimized sub-graph |
Matteo Martincigh | 8492433 | 2019-05-09 12:46:16 +0100 | [diff] [blame] | 550 | for (auto& substitution : optimizationViews.GetSubstitutions()) |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 551 | { |
| 552 | // Sub-graph optimized, substitute the sub-graph with the new optimized one in the main optimized graph |
Matteo Martincigh | 8492433 | 2019-05-09 12:46:16 +0100 | [diff] [blame] | 553 | SubgraphView& replacementSubgraph = substitution.m_ReplacementSubgraph; |
| 554 | SubgraphView& substitutableSubgraph = substitution.m_SubstitutableSubgraph; |
| 555 | optGraph.SubstituteSubgraph(substitutableSubgraph, replacementSubgraph); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 556 | |
| 557 | // Assign the current backend to the optimized sub-graph |
Matteo Martincigh | 8492433 | 2019-05-09 12:46:16 +0100 | [diff] [blame] | 558 | std::for_each(replacementSubgraph.begin(), replacementSubgraph.end(), [&selectedBackend](Layer* l) |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 559 | { |
| 560 | BOOST_ASSERT(l); |
| 561 | l->SetBackendId(selectedBackend); |
| 562 | }); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 563 | } |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 564 | |
Matteo Martincigh | 8492433 | 2019-05-09 12:46:16 +0100 | [diff] [blame] | 565 | if (!optimizationViews.GetFailedSubgraphs().empty()) |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 566 | { |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 567 | std::stringstream warningMsg; |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 568 | warningMsg << "Some sub-graph(s) failed to optimized on " << backendObjPtr->GetId() << " backend."; |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 569 | ReportWarning(warningMsg.str(), errMessages); |
| 570 | |
| 571 | // Failed to optimize the given sub-graph, re-assign the sub-graph layers to other available backends |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 572 | BackendSettings settingsCopy(backendSettings); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 573 | if (!backendObjPtr->GetId().IsCpuRef()) |
| 574 | { |
| 575 | // Add the current backend to the list of backends to ignore |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 576 | settingsCopy.m_IgnoredBackends.insert(backendObjPtr->GetId()); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 577 | } |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 578 | |
| 579 | int count=0; |
Matteo Martincigh | 8492433 | 2019-05-09 12:46:16 +0100 | [diff] [blame] | 580 | for (auto& failedSubgraph : optimizationViews.GetFailedSubgraphs()) |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 581 | { |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 582 | // An error occurred: the optimization was attempted but not performed, try different backends |
| 583 | std::stringstream subgraphMsg; |
| 584 | subgraphMsg << "Re-assigning backends to " << failedSubgraph.GetLayers().size() |
| 585 | << " layers inside sub-graph " << count++; |
Matteo Martincigh | 328d92b | 2019-07-04 17:52:55 +0100 | [diff] [blame] | 586 | ReportWarning(subgraphMsg.str(), errMessages); |
Derek Lamberti | c2fe5fb | 2019-05-08 10:23:08 +0100 | [diff] [blame] | 587 | |
| 588 | OptimizationResult reassignmentResult = AssignBackends(optNetObjPtr, |
| 589 | settingsCopy, |
| 590 | *subgraph, |
| 591 | errMessages); |
| 592 | if (reassignmentResult.m_Error) |
| 593 | { |
| 594 | // Failed to re-assign one of the remaining backends to each layer of the sub-graph |
| 595 | result.m_Error = true; |
| 596 | return result; |
| 597 | } |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 598 | } |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 599 | } |
| 600 | } |
| 601 | } |
| 602 | |
| 603 | return result; |
| 604 | } |
| 605 | |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 606 | bool RequiresCopy(ITensorHandleFactory::FactoryId src, |
| 607 | ITensorHandleFactory::FactoryId dst, |
| 608 | TensorHandleFactoryRegistry& registry) |
| 609 | { |
| 610 | if (src != dst) |
| 611 | { |
| 612 | ITensorHandleFactory* srcFactory = registry.GetFactory(src); |
| 613 | ITensorHandleFactory* dstFactory = registry.GetFactory(dst); |
| 614 | |
Matteo Martincigh | a6539ed | 2019-08-27 13:43:32 +0100 | [diff] [blame] | 615 | if (srcFactory && dstFactory && |
| 616 | (srcFactory->GetExportFlags() & dstFactory->GetImportFlags()) != 0) |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 617 | { |
| 618 | return false; |
| 619 | } |
| 620 | return true; |
| 621 | } |
| 622 | return false; |
| 623 | } |
| 624 | |
| 625 | // Find the handle factory for the input layer which results in fewest required copies. |
| 626 | ITensorHandleFactory::FactoryId CalculateSlotOptionForInput(BackendsMap& backends, |
| 627 | OutputSlot& slot, |
| 628 | TensorHandleFactoryRegistry& registry) |
| 629 | { |
| 630 | Layer& layer = slot.GetOwningLayer(); |
| 631 | BOOST_ASSERT(layer.GetType() == LayerType::Input); |
| 632 | |
| 633 | // Explicitly select the tensorhandle factory for InputLayer because the rules for it are slightly different. It |
| 634 | // doesn't matter which backend it is assigned to because they all use the same implementation, which |
| 635 | // requires Map/Unmap support. This means that, so long as the handle type supports map/unmap semantics, we can |
| 636 | // select a factory with maximum compatibility with the layers connected to the InputLayer. |
| 637 | |
| 638 | // First ensure the from backends can support the TensorHandeAPI |
| 639 | auto frmBackend = backends.find(layer.GetBackendId()); |
| 640 | if (frmBackend == backends.end() || |
| 641 | !frmBackend->second->SupportsTensorAllocatorAPI()) |
| 642 | { |
| 643 | return ITensorHandleFactory::LegacyFactoryId; |
| 644 | } |
| 645 | |
| 646 | // Go through all connections to the output slot and determine the TensorHandleFactory which results in the |
| 647 | // fewest copies. |
| 648 | std::map<ITensorHandleFactory::FactoryId, int> factoryScores; |
| 649 | int topScore = 0; |
| 650 | ITensorHandleFactory::FactoryId topChoice = ITensorHandleFactory::LegacyFactoryId; |
| 651 | |
| 652 | for (auto&& connection : slot.GetConnections()) |
| 653 | { |
| 654 | const Layer& connectedLayer = connection->GetOwningLayer(); |
| 655 | |
| 656 | auto toBackend = backends.find(connectedLayer.GetBackendId()); |
| 657 | BOOST_ASSERT_MSG(toBackend != backends.end(), "Backend id not found for the connected layer"); |
| 658 | |
| 659 | if (!toBackend->second.get()->SupportsTensorAllocatorAPI()) |
| 660 | { |
| 661 | // The destination backend does not support the tensor allocator API, move to the next one |
| 662 | continue; |
| 663 | } |
| 664 | |
| 665 | auto dstPrefs = toBackend->second.get()->GetHandleFactoryPreferences(); |
| 666 | for (auto&& dst : dstPrefs) |
| 667 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 668 | // Input layers use the mem copy workload or import, so the selected factory must |
| 669 | // support either the map/unmap API or Import API |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 670 | ITensorHandleFactory* factory = registry.GetFactory(dst); |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 671 | if (!factory->SupportsMapUnmap() && |
| 672 | !CheckFlag(factory->GetImportFlags(), MemorySource::Malloc)) // Just support cpu mem imports for now |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 673 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 674 | // The current tensor handle factory does not support the map/unmap or import |
| 675 | // strategy, move to the next one |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 676 | continue; |
| 677 | } |
| 678 | |
| 679 | auto it = factoryScores.find(dst); |
| 680 | if (it == factoryScores.end()) |
| 681 | { |
| 682 | // Add new score to the table |
| 683 | factoryScores[dst] = 0; |
| 684 | if (topChoice == ITensorHandleFactory::LegacyFactoryId) |
| 685 | { |
| 686 | topChoice = dst; |
| 687 | } |
| 688 | } |
| 689 | else |
| 690 | { |
| 691 | // Increase the score |
| 692 | factoryScores[dst]++; |
| 693 | |
| 694 | // Track the best option |
| 695 | if (factoryScores[dst] > topScore) |
| 696 | { |
| 697 | topScore = factoryScores[dst]; |
| 698 | topChoice = dst; |
| 699 | } |
| 700 | } |
| 701 | } |
| 702 | } |
| 703 | |
| 704 | return topChoice; |
| 705 | } |
| 706 | |
| 707 | // Find the handle factory for the output layer which results in fewest required copies. |
| 708 | ITensorHandleFactory::FactoryId CalculateSlotOptionForOutput(BackendsMap& backends, |
| 709 | OutputSlot& slot, |
| 710 | TensorHandleFactoryRegistry& registry) |
| 711 | { |
Jan Eilers | 8eb2560 | 2020-03-09 12:13:48 +0000 | [diff] [blame] | 712 | IgnoreUnused(backends, slot, registry); |
Derek Lamberti | 94a88d2 | 2019-12-10 21:12:59 +0000 | [diff] [blame] | 713 | return ITensorHandleFactory::DeferredFactoryId; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 714 | } |
| 715 | |
| 716 | // For all handle factories supported on the source backend, we wish to find the one which requires the fewest copies |
| 717 | // when considering all connections. |
| 718 | ITensorHandleFactory::FactoryId CalculateSlotOption(BackendsMap& backends, |
| 719 | OutputSlot& outputSlot, |
| 720 | TensorHandleFactoryRegistry& registry) |
| 721 | { |
| 722 | // First ensure the from backends can support the TensorHandeAPI |
| 723 | Layer& layer = outputSlot.GetOwningLayer(); |
| 724 | auto frmBackend = backends.find(layer.GetBackendId()); |
| 725 | if (frmBackend == backends.end() || |
| 726 | !frmBackend->second->SupportsTensorAllocatorAPI()) |
| 727 | { |
| 728 | return ITensorHandleFactory::LegacyFactoryId; |
| 729 | } |
| 730 | |
| 731 | // Connections to Output Layers requires support for map/unmap on the TensorHandle. |
| 732 | bool requiresMapUnmap = false; |
| 733 | for (auto&& connection : outputSlot.GetConnections()) |
| 734 | { |
| 735 | const Layer& connectedLayer = connection->GetOwningLayer(); |
| 736 | if (connectedLayer.GetType() == LayerType::Output) |
| 737 | { |
| 738 | requiresMapUnmap = true; |
| 739 | } |
| 740 | } |
| 741 | |
| 742 | IBackendInternal* srcBackend = frmBackend->second.get(); |
| 743 | auto srcPrefs = srcBackend->GetHandleFactoryPreferences(); |
| 744 | |
| 745 | // Initialize the scores |
| 746 | std::map<ITensorHandleFactory::FactoryId, int> factoryScores; |
| 747 | for (auto&& pref : srcPrefs) |
| 748 | { |
| 749 | if (requiresMapUnmap) // Only consider factories that support map/unmap if required |
| 750 | { |
| 751 | ITensorHandleFactory* factory = registry.GetFactory(pref); |
| 752 | if (!factory->SupportsMapUnmap()) |
| 753 | { |
| 754 | // The current tensor handle factory does not support the map/unmap strategy, move to the next one |
| 755 | continue; |
| 756 | } |
| 757 | } |
| 758 | |
| 759 | auto it = factoryScores.find(pref); |
| 760 | if (it == factoryScores.end()) |
| 761 | { |
| 762 | // Add new score to the table |
| 763 | factoryScores[pref] = 0; |
| 764 | } |
| 765 | } |
| 766 | |
| 767 | // Score each handle factory based on how many times it requires copies on the slot connections |
| 768 | for (auto&& connection : outputSlot.GetConnections()) |
| 769 | { |
| 770 | const Layer& connectedLayer = connection->GetOwningLayer(); |
| 771 | |
| 772 | auto toBackend = backends.find(connectedLayer.GetBackendId()); |
| 773 | BOOST_ASSERT_MSG(toBackend != backends.end(), "Backend id not found for the connected layer"); |
| 774 | |
| 775 | auto dstPrefs = toBackend->second.get()->GetHandleFactoryPreferences(); |
| 776 | for (auto&& src : srcPrefs) |
| 777 | { |
| 778 | if (factoryScores.find(src) == factoryScores.end()) // Don't consider excluded factories |
| 779 | { |
| 780 | continue; |
| 781 | } |
| 782 | |
| 783 | for (auto&& dst : dstPrefs) |
| 784 | { |
| 785 | if (RequiresCopy(src, dst, registry)) |
| 786 | { |
| 787 | // Copy avoided, increase the score |
| 788 | factoryScores[src]++; |
| 789 | break; |
| 790 | } |
| 791 | } |
| 792 | } |
| 793 | } |
| 794 | |
| 795 | // Find the lowest score |
| 796 | int minScore = std::numeric_limits<int>::max(); |
| 797 | for (auto it : factoryScores) |
| 798 | { |
| 799 | minScore = std::min(minScore, it.second); |
| 800 | } |
| 801 | |
| 802 | // Collect factories matching the best(lowest) score |
| 803 | std::vector<ITensorHandleFactory::FactoryId> optimalFactories; |
| 804 | for (auto it : factoryScores) |
| 805 | { |
| 806 | if (it.second == minScore) |
| 807 | { |
| 808 | optimalFactories.push_back(it.first); |
| 809 | } |
| 810 | } |
| 811 | |
| 812 | // For all compatible Factories matching the best score, find the preferred one for the current layer. |
| 813 | for (auto&& srcPref : srcPrefs) |
| 814 | { |
| 815 | for (auto&& comp : optimalFactories) |
| 816 | { |
| 817 | if (comp == srcPref) |
| 818 | { |
| 819 | return comp; |
| 820 | } |
| 821 | } |
| 822 | } |
| 823 | |
| 824 | return ITensorHandleFactory::LegacyFactoryId; |
| 825 | } |
| 826 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 827 | EdgeStrategy CalculateEdgeStrategy(BackendsMap& backends, |
| 828 | ITensorHandleFactory::FactoryId srcFactoryId, |
| 829 | const Layer& layer, |
| 830 | const Layer& connectedLayer, |
| 831 | TensorHandleFactoryRegistry& registry) |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 832 | { |
| 833 | auto toBackend = backends.find(connectedLayer.GetBackendId()); |
| 834 | BOOST_ASSERT_MSG(toBackend != backends.end(), "Backend id not found for the connected layer"); |
| 835 | |
| 836 | auto dstPrefs = toBackend->second.get()->GetHandleFactoryPreferences(); |
| 837 | |
| 838 | // Legacy API check for backward compatibility |
| 839 | if (srcFactoryId == ITensorHandleFactory::LegacyFactoryId || dstPrefs.empty()) |
| 840 | { |
| 841 | if (layer.GetBackendId() != connectedLayer.GetBackendId()) |
| 842 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 843 | return EdgeStrategy::CopyToTarget; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 844 | } |
| 845 | else |
| 846 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 847 | return EdgeStrategy::DirectCompatibility; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 848 | } |
| 849 | } |
| 850 | |
| 851 | // TensorHandleFactory API present, so perform more sophisticated strategies. |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 852 | // Dst Output layers don't require copy because they use import or map/unmap |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 853 | if (connectedLayer.GetType() == LayerType::Output) |
| 854 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 855 | return EdgeStrategy::DirectCompatibility; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 856 | } |
| 857 | |
| 858 | // Search for direct match in prefs |
| 859 | for (auto&& pref : dstPrefs) |
| 860 | { |
| 861 | if (pref == srcFactoryId) |
| 862 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 863 | return EdgeStrategy::DirectCompatibility; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 864 | } |
| 865 | } |
| 866 | |
| 867 | // Search for export/import options |
| 868 | ITensorHandleFactory* srcFactory = registry.GetFactory(srcFactoryId); |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 869 | if (srcFactory->GetExportFlags() != 0) |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 870 | { |
| 871 | for (auto&& pref : dstPrefs) |
| 872 | { |
| 873 | ITensorHandleFactory* dstFactory = registry.GetFactory(pref); |
James Conroy | ffab16f | 2019-11-07 14:37:09 +0000 | [diff] [blame] | 874 | |
James Conroy | 47e863d | 2019-11-18 17:07:43 +0000 | [diff] [blame] | 875 | // Handles cases when a destPref is not listed in TensorHandleFactoryRegistry |
James Conroy | ffab16f | 2019-11-07 14:37:09 +0000 | [diff] [blame] | 876 | if (!dstFactory) { |
James Conroy | 47e863d | 2019-11-18 17:07:43 +0000 | [diff] [blame] | 877 | continue; |
James Conroy | ffab16f | 2019-11-07 14:37:09 +0000 | [diff] [blame] | 878 | } |
| 879 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 880 | if ((dstFactory->GetImportFlags() & srcFactory->GetExportFlags()) != 0) |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 881 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 882 | return EdgeStrategy::ExportToTarget; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 883 | } |
| 884 | } |
| 885 | } |
| 886 | |
| 887 | // Search for copy options via map/unmap |
| 888 | if (srcFactory->SupportsMapUnmap()) |
| 889 | { |
| 890 | for (auto&& pref : dstPrefs) |
| 891 | { |
| 892 | ITensorHandleFactory* dstFactory = registry.GetFactory(pref); |
James Conroy | 47e863d | 2019-11-18 17:07:43 +0000 | [diff] [blame] | 893 | if (dstFactory && dstFactory->SupportsMapUnmap()) |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 894 | { |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 895 | return EdgeStrategy::CopyToTarget; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 896 | } |
| 897 | } |
| 898 | } |
| 899 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 900 | return EdgeStrategy::Undefined; |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 901 | } |
| 902 | |
| 903 | // Select the TensorHandleFactories and the corresponding memory strategy |
| 904 | OptimizationResult SelectTensorHandleStrategy(Graph& optGraph, |
| 905 | BackendsMap& backends, |
| 906 | TensorHandleFactoryRegistry& registry, |
| 907 | Optional<std::vector<std::string>&> errMessages) |
| 908 | { |
| 909 | OptimizationResult result; |
| 910 | |
| 911 | optGraph.ForEachLayer([&backends, ®istry, &result, &errMessages](Layer* layer) |
| 912 | { |
| 913 | BOOST_ASSERT(layer); |
| 914 | |
| 915 | // Lets make sure the backend is in our list of supported backends. Something went wrong during backend |
| 916 | // assignment if this check fails |
| 917 | BOOST_ASSERT(backends.find(layer->GetBackendId()) != backends.end()); |
| 918 | |
| 919 | // Check each output separately |
| 920 | for (unsigned int slotIdx = 0; slotIdx < layer->GetNumOutputSlots(); slotIdx++) |
| 921 | { |
| 922 | OutputSlot& outputSlot = layer->GetOutputSlot(slotIdx); |
| 923 | |
| 924 | ITensorHandleFactory::FactoryId slotOption = ITensorHandleFactory::LegacyFactoryId; |
| 925 | |
| 926 | // Calculate the factory to use which results in the fewest copies being made. |
| 927 | switch(layer->GetType()) |
| 928 | { |
| 929 | case LayerType::Input: |
| 930 | slotOption = CalculateSlotOptionForInput(backends, outputSlot, registry); |
| 931 | break; |
| 932 | case LayerType::Output: |
| 933 | slotOption = CalculateSlotOptionForOutput(backends, outputSlot, registry); |
| 934 | break; |
| 935 | default: |
| 936 | slotOption = CalculateSlotOption(backends, outputSlot, registry); |
| 937 | break; |
| 938 | } |
| 939 | outputSlot.SetTensorHandleFactory(slotOption); |
| 940 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 941 | // Now determine the "best" edge strategy for each connection given the slotOption. |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 942 | unsigned int connectionIdx = 0; |
| 943 | for (auto&& connection : outputSlot.GetConnections()) |
| 944 | { |
| 945 | const Layer& connectedLayer = connection->GetOwningLayer(); |
| 946 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 947 | EdgeStrategy strategy = CalculateEdgeStrategy(backends, slotOption, *layer, connectedLayer, registry); |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 948 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 949 | if (strategy == EdgeStrategy::Undefined) |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 950 | { |
| 951 | result.m_Error = true; |
| 952 | if (errMessages) |
| 953 | { |
| 954 | errMessages.value().emplace_back("Could not find valid strategy required for compatibility" |
| 955 | " between backends."); |
| 956 | } |
| 957 | return; |
| 958 | } |
| 959 | |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 960 | outputSlot.SetEdgeStrategy(connectionIdx, strategy); |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 961 | |
| 962 | connectionIdx++; |
| 963 | } |
| 964 | } |
| 965 | }); |
| 966 | |
| 967 | return result; |
| 968 | } |
| 969 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 970 | IOptimizedNetworkPtr Optimize(const INetwork& inNetwork, |
| 971 | const std::vector<BackendId>& backendPreferences, |
| 972 | const IDeviceSpec& deviceSpec, |
| 973 | const OptimizerOptions& options, |
Rob Hughes | 2321443 | 2019-11-05 11:27:36 +0000 | [diff] [blame] | 974 | Optional<std::vector<std::string>&> messages) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 975 | { |
| 976 | if (backendPreferences.empty()) |
| 977 | { |
| 978 | throw armnn::InvalidArgumentException("Invoked Optimize with no backends specified"); |
| 979 | } |
| 980 | |
Narumol Prangnawarat | bc7ffb5 | 2020-03-20 15:01:01 +0000 | [diff] [blame^] | 981 | if (options.m_ReduceFp32ToFp16 && options.m_ReduceFp32ToBf16) |
| 982 | { |
| 983 | throw InvalidArgumentException("BFloat16 and Float16 optimization cannot be enabled at the same time."); |
| 984 | } |
| 985 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 986 | const Network& network = *boost::polymorphic_downcast<const Network*>(&inNetwork); |
| 987 | std::unique_ptr<Graph> graph = std::make_unique<Graph>(network.GetGraph()); |
| 988 | |
| 989 | auto optNet = IOptimizedNetworkPtr(new OptimizedNetwork(std::move(graph)), &IOptimizedNetwork::Destroy); |
| 990 | |
| 991 | OptimizedNetwork* optNetObjPtr = boost::polymorphic_downcast<OptimizedNetwork*>(optNet.get()); |
| 992 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 993 | // Get the optimized graph |
| 994 | Graph& optGraph = optNetObjPtr->GetGraph(); |
| 995 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 996 | // Perform optimisation passes |
| 997 | using namespace optimizations; |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 998 | Optimizer::Pass(optGraph, MakeOptimizations(SquashEqualPermuteSiblings(), |
Mike Kelly | 490b7be | 2020-03-03 12:39:09 +0000 | [diff] [blame] | 999 | SquashEqualTransposeSiblings(), |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1000 | SquashEqualReshapeSiblings(), |
| 1001 | OptimizeInversePermutes(), |
Mike Kelly | 490b7be | 2020-03-03 12:39:09 +0000 | [diff] [blame] | 1002 | OptimizeInverseTransposes(), |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1003 | MovePermuteUp(), |
Mike Kelly | 490b7be | 2020-03-03 12:39:09 +0000 | [diff] [blame] | 1004 | MoveTransposeUp(), |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1005 | PermuteAsReshape(), |
Mike Kelly | 490b7be | 2020-03-03 12:39:09 +0000 | [diff] [blame] | 1006 | TransposeAsReshape(), |
Nina Drozd | 861985f | 2019-04-18 14:48:51 +0100 | [diff] [blame] | 1007 | OptimizeConsecutiveReshapes(), |
Rob Hughes | 3a7d3a7 | 2019-09-24 16:59:56 +0100 | [diff] [blame] | 1008 | FoldPadIntoConvolution2d(), |
Mike Kelly | 490b7be | 2020-03-03 12:39:09 +0000 | [diff] [blame] | 1009 | PermuteAndBatchToSpaceAsDepthToSpace(), |
| 1010 | TransposeAndBatchToSpaceAsDepthToSpace())); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1011 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1012 | // Infer the tensor infos for all output slots. Throws an exception on failure |
| 1013 | optGraph.InferTensorInfos(); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1014 | |
| 1015 | // If Fp32 to Fp16 optimization is set convert Fp32 network to Fp16 |
| 1016 | if (options.m_ReduceFp32ToFp16) |
| 1017 | { |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1018 | Optimizer::Pass(optGraph, MakeOptimizations(Fp32NetworkToFp16Converter())); |
Derek Lamberti | dd6804b | 2019-11-27 09:29:57 +0000 | [diff] [blame] | 1019 | Optimizer::Pass(optGraph, MakeOptimizations(ConvertConstantsFloatToHalf())); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1020 | } |
| 1021 | |
Narumol Prangnawarat | bc7ffb5 | 2020-03-20 15:01:01 +0000 | [diff] [blame^] | 1022 | // If Fp32 to Bf16 optimization is set convert Fp32 network to Bf16 |
| 1023 | if (options.m_ReduceFp32ToBf16) |
| 1024 | { |
| 1025 | Optimizer::Pass(optGraph, MakeOptimizations(Fp32NetworkToBf16Converter())); |
| 1026 | Optimizer::Pass(optGraph, MakeOptimizations(ConvertConstantsFloatToBFloat())); |
| 1027 | } |
| 1028 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1029 | // Initialize backend settings |
| 1030 | BackendSettings backendSettings(backendPreferences, deviceSpec); |
| 1031 | if (backendSettings.GetAvailablePreferredBackends().empty()) |
| 1032 | { |
| 1033 | std::stringstream failureMsg; |
| 1034 | failureMsg << "None of the preferred backends " << backendPreferences |
| 1035 | << " are supported. Current platform provides " << backendSettings.m_SupportedBackends; |
Rob Hughes | 2321443 | 2019-11-05 11:27:36 +0000 | [diff] [blame] | 1036 | ReportError(failureMsg.str(), messages); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1037 | return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy); |
| 1038 | } |
| 1039 | |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1040 | // Create a map to temporarily hold initialized backend objects |
| 1041 | TensorHandleFactoryRegistry tensorHandleFactoryRegistry; |
| 1042 | BackendsMap backends = CreateSupportedBackends(tensorHandleFactoryRegistry, backendSettings); |
| 1043 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1044 | // Assign an available backend to each layer |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1045 | Graph::Iterator firstLayer = optGraph.begin(); |
| 1046 | Graph::Iterator lastLayer = optGraph.end(); |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1047 | OptimizationResult assignBackendsResult = AssignBackends(optNetObjPtr, |
| 1048 | backendSettings, |
| 1049 | firstLayer, |
| 1050 | lastLayer, |
Rob Hughes | 2321443 | 2019-11-05 11:27:36 +0000 | [diff] [blame] | 1051 | messages); |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1052 | if (assignBackendsResult.m_Error) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1053 | { |
| 1054 | // Failed to assign a backend to each layer |
jimfly01 | 6b0b53d | 2018-10-08 14:43:01 +0100 | [diff] [blame] | 1055 | return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy); |
| 1056 | } |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1057 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1058 | Optimizer::Pass(optGraph, MakeOptimizations(OptimizeInverseConversionsFp16(), |
| 1059 | OptimizeInverseConversionsFp32())); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1060 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1061 | // Apply the backend-specific optimizations |
| 1062 | OptimizationResult backendOptimizationResult = ApplyBackendOptimizations(optNetObjPtr, |
| 1063 | backendSettings, |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1064 | backends, |
Rob Hughes | 2321443 | 2019-11-05 11:27:36 +0000 | [diff] [blame] | 1065 | messages); |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1066 | if (backendOptimizationResult.m_Error) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1067 | { |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1068 | // Failed to apply the backend-specific optimizations |
| 1069 | return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1070 | } |
| 1071 | |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1072 | // If the debug flag is set, then insert a DebugLayer after each layer |
| 1073 | // Doing this after applying the backend optimizations as they might have changed some layers |
| 1074 | if (options.m_Debug) |
| 1075 | { |
| 1076 | Optimizer::Pass(optGraph, MakeOptimizations(InsertDebugLayer())); |
| 1077 | } |
| 1078 | |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1079 | // Calculate the compatibility strategies for tensor handles |
| 1080 | OptimizationResult strategyResult = SelectTensorHandleStrategy(optGraph, |
| 1081 | backends, |
| 1082 | tensorHandleFactoryRegistry, |
Rob Hughes | 2321443 | 2019-11-05 11:27:36 +0000 | [diff] [blame] | 1083 | messages); |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1084 | if (strategyResult.m_Error) |
| 1085 | { |
| 1086 | // Failed to apply the backend-specific optimizations |
| 1087 | return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy); |
| 1088 | } |
| 1089 | |
| 1090 | // Based on the tensor handle strategy determined above, insert copy layers where required. |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 1091 | optGraph.AddCompatibilityLayers(backends, tensorHandleFactoryRegistry); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1092 | |
| 1093 | // Convert constants |
Matteo Martincigh | adddddb | 2019-01-24 14:06:23 +0000 | [diff] [blame] | 1094 | Optimizer::Pass(optGraph, MakeOptimizations(ConvertConstantsFloatToHalf())); |
| 1095 | Optimizer::Pass(optGraph, MakeOptimizations(ConvertConstantsHalfToFloat())); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1096 | |
Derek Lamberti | 84da38b | 2019-06-13 11:40:08 +0100 | [diff] [blame] | 1097 | // Run backend specific optimizations (deprecated) |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1098 | for (auto&& chosenBackend : backendSettings.m_SelectedBackends) |
David Beck | 263e349 | 2018-11-09 14:46:40 +0000 | [diff] [blame] | 1099 | { |
| 1100 | auto factoryFun = BackendRegistryInstance().GetFactory(chosenBackend); |
| 1101 | auto backendPtr = factoryFun(); |
| 1102 | BOOST_ASSERT(backendPtr.get() != nullptr); |
| 1103 | |
Matteo Martincigh | ed73504 | 2019-05-22 09:42:43 +0100 | [diff] [blame] | 1104 | ARMNN_NO_DEPRECATE_WARN_BEGIN |
David Beck | 263e349 | 2018-11-09 14:46:40 +0000 | [diff] [blame] | 1105 | auto backendSpecificOptimizations = backendPtr->GetOptimizations(); |
Matteo Martincigh | ed73504 | 2019-05-22 09:42:43 +0100 | [diff] [blame] | 1106 | ARMNN_NO_DEPRECATE_WARN_END |
| 1107 | |
David Beck | 263e349 | 2018-11-09 14:46:40 +0000 | [diff] [blame] | 1108 | if (!backendSpecificOptimizations.empty()) |
| 1109 | { |
| 1110 | Optimizer::Pass(optNetObjPtr->GetGraph(), backendSpecificOptimizations); |
| 1111 | } |
| 1112 | } |
| 1113 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1114 | return optNet; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1115 | } |
| 1116 | |
| 1117 | Network::Network() |
Sadik Armagan | 3184c90 | 2020-03-18 10:57:30 +0000 | [diff] [blame] | 1118 | : m_Graph(std::make_unique<Graph>()) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1119 | { |
| 1120 | } |
| 1121 | |
| 1122 | Network::~Network() |
| 1123 | { |
| 1124 | } |
| 1125 | |
Jan Eilers | 99d9d4a | 2019-11-06 10:02:16 +0000 | [diff] [blame] | 1126 | Status Network::PrintGraph() |
| 1127 | { |
| 1128 | m_Graph->Print(); |
| 1129 | return Status::Success; |
| 1130 | } |
| 1131 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1132 | IConnectableLayer* Network::AddInputLayer(LayerBindingId id, const char* name) |
| 1133 | { |
| 1134 | return m_Graph->AddLayer<InputLayer>(id, name); |
| 1135 | } |
| 1136 | |
Éanna Ó Catháin | 4e1e136 | 2018-11-12 11:36:34 +0000 | [diff] [blame] | 1137 | IConnectableLayer* Network::AddBatchToSpaceNdLayer(const BatchToSpaceNdDescriptor& batchToSpaceNdDescriptor, |
| 1138 | const char* name) |
| 1139 | { |
| 1140 | return m_Graph->AddLayer<BatchToSpaceNdLayer>(batchToSpaceNdDescriptor, name); |
| 1141 | } |
| 1142 | |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 1143 | IConnectableLayer* Network::AddComparisonLayer(const ComparisonDescriptor& comparisonDescriptor, |
| 1144 | const char* name) |
| 1145 | { |
| 1146 | return m_Graph->AddLayer<ComparisonLayer>(comparisonDescriptor, name); |
| 1147 | } |
| 1148 | |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 1149 | IConnectableLayer* Network::AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor& elementwiseUnaryDescriptor, |
| 1150 | const char* name) |
| 1151 | { |
| 1152 | return m_Graph->AddLayer<ElementwiseUnaryLayer>(elementwiseUnaryDescriptor, name); |
| 1153 | } |
| 1154 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1155 | IConnectableLayer* Network::AddFullyConnectedLayerImpl(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1156 | const ConstTensor& weights, |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1157 | const Optional<ConstTensor>& biases, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1158 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1159 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1160 | if (fullyConnectedDescriptor.m_BiasEnabled && !biases.has_value()) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1161 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1162 | throw InvalidArgumentException("AddFullyConnectedLayer: biases cannot be empty"); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1163 | } |
| 1164 | |
| 1165 | const auto layer = m_Graph->AddLayer<FullyConnectedLayer>(fullyConnectedDescriptor, name); |
| 1166 | |
| 1167 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 1168 | |
| 1169 | if (fullyConnectedDescriptor.m_BiasEnabled) |
| 1170 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1171 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(biases.value()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1172 | } |
| 1173 | |
| 1174 | return layer; |
| 1175 | } |
| 1176 | |
| 1177 | IConnectableLayer* Network::AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1178 | const ConstTensor& weights, |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1179 | const Optional<ConstTensor>& biases, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1180 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1181 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1182 | return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, weights, biases, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1183 | } |
| 1184 | |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1185 | IConnectableLayer* Network::AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
| 1186 | const ConstTensor& weights, |
| 1187 | const char* name) |
| 1188 | { |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1189 | Optional<ConstTensor> biases; |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1190 | return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, weights, biases, name); |
| 1191 | } |
| 1192 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1193 | IConnectableLayer* Network::AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1194 | const ConstTensor& weights, |
| 1195 | const ConstTensor& biases, |
| 1196 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1197 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1198 | Optional<ConstTensor> optionalBiases(biases); |
| 1199 | return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, weights, optionalBiases, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1200 | } |
| 1201 | |
Jim Flynn | e242f2d | 2019-05-22 14:24:13 +0100 | [diff] [blame] | 1202 | IConnectableLayer* Network::AddConcatLayer(const ConcatDescriptor& concatDescriptor, |
Jim Flynn | 906f946 | 2019-05-10 13:55:21 +0100 | [diff] [blame] | 1203 | const char* name) |
| 1204 | { |
Jim Flynn | e242f2d | 2019-05-22 14:24:13 +0100 | [diff] [blame] | 1205 | return m_Graph->AddLayer<ConcatLayer>(concatDescriptor, name); |
Jim Flynn | 906f946 | 2019-05-10 13:55:21 +0100 | [diff] [blame] | 1206 | } |
| 1207 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1208 | IConnectableLayer* Network::AddConvolution2dLayerImpl(const Convolution2dDescriptor& convolution2dDescriptor, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1209 | const ConstTensor& weights, |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1210 | const Optional<ConstTensor>& biases, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1211 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1212 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1213 | if (convolution2dDescriptor.m_BiasEnabled && !biases.has_value()) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1214 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1215 | throw InvalidArgumentException("AddConvolution2dLayer: biases cannot be empty"); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1216 | } |
| 1217 | |
| 1218 | const auto layer = m_Graph->AddLayer<Convolution2dLayer>(convolution2dDescriptor, name); |
| 1219 | |
| 1220 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 1221 | |
| 1222 | if (convolution2dDescriptor.m_BiasEnabled) |
| 1223 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1224 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(biases.value()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1225 | } |
| 1226 | |
| 1227 | return layer; |
| 1228 | } |
| 1229 | |
| 1230 | IConnectableLayer* Network::AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1231 | const ConstTensor& weights, |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1232 | const Optional<ConstTensor>& biases, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1233 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1234 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1235 | return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1236 | } |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1237 | |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1238 | IConnectableLayer* Network::AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, |
| 1239 | const ConstTensor& weights, |
| 1240 | const char* name) |
| 1241 | { |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1242 | Optional<ConstTensor> biases; |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1243 | return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name); |
| 1244 | } |
| 1245 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1246 | IConnectableLayer* Network::AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1247 | const ConstTensor& weights, |
| 1248 | const ConstTensor& biases, |
| 1249 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1250 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1251 | Optional<ConstTensor> optionalBiases(biases); |
| 1252 | return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, optionalBiases, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1253 | } |
| 1254 | |
| 1255 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayerImpl( |
| 1256 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 1257 | const ConstTensor& weights, |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1258 | const Optional<ConstTensor>& biases, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1259 | const char* name) |
| 1260 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1261 | if (convolution2dDescriptor.m_BiasEnabled && !biases.has_value()) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1262 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1263 | throw InvalidArgumentException("AddDepthwiseConvolution2dLayer: biases cannot be empty"); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1264 | } |
| 1265 | |
Matteo Martincigh | 3d6898c | 2019-01-15 16:11:44 +0000 | [diff] [blame] | 1266 | const auto layer = m_Graph->AddLayer<DepthwiseConvolution2dLayer>(convolution2dDescriptor, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1267 | |
| 1268 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 1269 | |
| 1270 | if (convolution2dDescriptor.m_BiasEnabled) |
| 1271 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1272 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(biases.value()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1273 | } |
| 1274 | |
| 1275 | return layer; |
| 1276 | } |
| 1277 | |
Aron Virginas-Tar | dd6247f | 2019-09-19 14:31:17 +0100 | [diff] [blame] | 1278 | IConnectableLayer* Network::AddDepthToSpaceLayer(const DepthToSpaceDescriptor& depthToSpaceDescriptor, |
| 1279 | const char* name) |
| 1280 | { |
| 1281 | return m_Graph->AddLayer<DepthToSpaceLayer>(depthToSpaceDescriptor, name); |
| 1282 | } |
| 1283 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1284 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayer( |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1285 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 1286 | const ConstTensor& weights, |
| 1287 | const Optional<ConstTensor>& biases, |
| 1288 | const char* name) |
| 1289 | { |
| 1290 | return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name); |
| 1291 | } |
| 1292 | |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1293 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayer( |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1294 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 1295 | const ConstTensor& weights, |
| 1296 | const char* name) |
| 1297 | { |
Matteo Martincigh | fc598e1 | 2019-05-14 10:36:13 +0100 | [diff] [blame] | 1298 | Optional<ConstTensor> biases; |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1299 | return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1300 | } |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1301 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1302 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayer( |
| 1303 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 1304 | const ConstTensor& weights, |
| 1305 | const ConstTensor& biases, |
| 1306 | const char* name) |
| 1307 | { |
Aron Virginas-Tar | ad40270 | 2019-02-22 17:03:44 +0000 | [diff] [blame] | 1308 | Optional<ConstTensor> optionalBiases(biases); |
| 1309 | return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, optionalBiases, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1310 | } |
| 1311 | |
Narumol Prangnawarat | 94dd5d8 | 2019-01-23 18:06:26 +0000 | [diff] [blame] | 1312 | IConnectableLayer* Network::AddDetectionPostProcessLayer(const armnn::DetectionPostProcessDescriptor& descriptor, |
Narumol Prangnawarat | 6d302bf | 2019-02-04 11:46:26 +0000 | [diff] [blame] | 1313 | const ConstTensor& anchors, const char* name) |
Narumol Prangnawarat | 94dd5d8 | 2019-01-23 18:06:26 +0000 | [diff] [blame] | 1314 | { |
Narumol Prangnawarat | 6d302bf | 2019-02-04 11:46:26 +0000 | [diff] [blame] | 1315 | const auto layer = m_Graph->AddLayer<DetectionPostProcessLayer>(descriptor, name); |
| 1316 | |
| 1317 | layer->m_Anchors = std::make_unique<ScopedCpuTensorHandle>(anchors); |
| 1318 | |
| 1319 | return layer; |
Narumol Prangnawarat | 94dd5d8 | 2019-01-23 18:06:26 +0000 | [diff] [blame] | 1320 | } |
| 1321 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1322 | IConnectableLayer* Network::AddPermuteLayer(const PermuteDescriptor& permuteDescriptor, |
| 1323 | const char* name) |
| 1324 | { |
| 1325 | return m_Graph->AddLayer<PermuteLayer>(permuteDescriptor, name); |
| 1326 | } |
| 1327 | |
| 1328 | IConnectableLayer* Network::AddPooling2dLayer(const Pooling2dDescriptor& pooling2dDescriptor, |
| 1329 | const char* name) |
| 1330 | { |
| 1331 | return m_Graph->AddLayer<Pooling2dLayer>(pooling2dDescriptor, name); |
| 1332 | } |
| 1333 | |
| 1334 | IConnectableLayer* Network::AddActivationLayer(const ActivationDescriptor& activationDescriptor, |
| 1335 | const char* name) |
| 1336 | { |
| 1337 | return m_Graph->AddLayer<ActivationLayer>(activationDescriptor, name); |
| 1338 | } |
| 1339 | |
Nikhil Raj | ee391d5 | 2019-09-05 17:50:44 +0100 | [diff] [blame] | 1340 | IConnectableLayer* Network::AddArgMinMaxLayer(const ArgMinMaxDescriptor& argMinMaxDescriptor, |
| 1341 | const char* name) |
| 1342 | { |
| 1343 | return m_Graph->AddLayer<ArgMinMaxLayer>(argMinMaxDescriptor, name); |
| 1344 | } |
| 1345 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1346 | IConnectableLayer* Network::AddNormalizationLayer(const NormalizationDescriptor& |
| 1347 | normalizationDescriptor, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1348 | const char* name) |
| 1349 | { |
| 1350 | return m_Graph->AddLayer<NormalizationLayer>(normalizationDescriptor, name); |
| 1351 | } |
| 1352 | |
Aron Virginas-Tar | 636ab40 | 2019-09-16 14:27:45 +0100 | [diff] [blame] | 1353 | IConnectableLayer* Network::AddSliceLayer(const SliceDescriptor& sliceDescriptor, const char* name) |
| 1354 | { |
| 1355 | return m_Graph->AddLayer<SliceLayer>(sliceDescriptor, name); |
| 1356 | } |
| 1357 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1358 | IConnectableLayer* Network::AddSoftmaxLayer(const SoftmaxDescriptor& softmaxDescriptor, |
| 1359 | const char* name) |
| 1360 | { |
| 1361 | return m_Graph->AddLayer<SoftmaxLayer>(softmaxDescriptor, name); |
| 1362 | } |
| 1363 | |
| 1364 | IConnectableLayer* Network::AddSplitterLayer(const ViewsDescriptor& splitterDescriptor, |
| 1365 | const char* name) |
| 1366 | { |
| 1367 | return m_Graph->AddLayer<SplitterLayer>(splitterDescriptor, name); |
| 1368 | } |
| 1369 | |
Nattapat Chaimanowong | 5a4304a | 2018-11-28 10:44:37 +0000 | [diff] [blame] | 1370 | IConnectableLayer* Network::AddMaximumLayer(const char* name) |
| 1371 | { |
| 1372 | return m_Graph->AddLayer<MaximumLayer>(name); |
| 1373 | } |
| 1374 | |
Éanna Ó Catháin | 20e5880 | 2018-12-04 10:29:06 +0000 | [diff] [blame] | 1375 | IConnectableLayer* Network::AddMinimumLayer(const char* name) |
| 1376 | { |
| 1377 | return m_Graph->AddLayer<MinimumLayer>(name); |
| 1378 | } |
| 1379 | |
Jim Flynn | e242f2d | 2019-05-22 14:24:13 +0100 | [diff] [blame] | 1380 | IConnectableLayer* Network::AddMergerLayer(const MergerDescriptor& mergerDescriptor, |
Jim Flynn | 906f946 | 2019-05-10 13:55:21 +0100 | [diff] [blame] | 1381 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1382 | { |
Jim Flynn | e242f2d | 2019-05-22 14:24:13 +0100 | [diff] [blame] | 1383 | return AddConcatLayer(mergerDescriptor, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1384 | } |
| 1385 | |
Kevin May | 868eb14 | 2019-09-04 17:29:31 +0100 | [diff] [blame] | 1386 | IConnectableLayer* Network::AddAbsLayer(const char * name) |
| 1387 | { |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 1388 | return AddElementwiseUnaryLayer(ElementwiseUnaryDescriptor(UnaryOperation::Abs), name); |
Kevin May | 868eb14 | 2019-09-04 17:29:31 +0100 | [diff] [blame] | 1389 | } |
| 1390 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1391 | IConnectableLayer* Network::AddAdditionLayer(const char* name) |
| 1392 | { |
| 1393 | return m_Graph->AddLayer<AdditionLayer>(name); |
| 1394 | } |
| 1395 | |
| 1396 | IConnectableLayer* Network::AddMultiplicationLayer(const char* name) |
| 1397 | { |
| 1398 | return m_Graph->AddLayer<MultiplicationLayer>(name); |
| 1399 | } |
| 1400 | |
| 1401 | IConnectableLayer* Network::AddOutputLayer(LayerBindingId id, const char* name) |
| 1402 | { |
| 1403 | return m_Graph->AddLayer<OutputLayer>(id, name); |
| 1404 | } |
| 1405 | |
| 1406 | IConnectableLayer* Network::AddBatchNormalizationLayer(const BatchNormalizationDescriptor& desc, |
| 1407 | const ConstTensor& mean, |
| 1408 | const ConstTensor& variance, |
| 1409 | const ConstTensor& beta, |
| 1410 | const ConstTensor& gamma, |
| 1411 | const char* name) |
| 1412 | { |
| 1413 | const auto layer = m_Graph->AddLayer<BatchNormalizationLayer>(desc, name); |
| 1414 | |
| 1415 | layer->m_Mean = std::make_unique<ScopedCpuTensorHandle>(mean); |
| 1416 | layer->m_Variance = std::make_unique<ScopedCpuTensorHandle>(variance); |
| 1417 | layer->m_Beta = std::make_unique<ScopedCpuTensorHandle>(beta); |
| 1418 | layer->m_Gamma = std::make_unique<ScopedCpuTensorHandle>(gamma); |
| 1419 | |
| 1420 | return layer; |
| 1421 | } |
| 1422 | |
Aron Virginas-Tar | 169d2f1 | 2019-07-01 19:01:44 +0100 | [diff] [blame] | 1423 | IConnectableLayer* Network::AddResizeBilinearLayer(const ResizeBilinearDescriptor& descriptor, |
| 1424 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1425 | { |
Aron Virginas-Tar | 169d2f1 | 2019-07-01 19:01:44 +0100 | [diff] [blame] | 1426 | ResizeDescriptor resizeDescriptor; |
| 1427 | resizeDescriptor.m_Method = ResizeMethod::Bilinear; |
| 1428 | resizeDescriptor.m_DataLayout = descriptor.m_DataLayout; |
| 1429 | resizeDescriptor.m_TargetWidth = descriptor.m_TargetWidth; |
| 1430 | resizeDescriptor.m_TargetHeight = descriptor.m_TargetHeight; |
| 1431 | |
| 1432 | return m_Graph->AddLayer<ResizeLayer>(resizeDescriptor, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1433 | } |
| 1434 | |
Teresa Charlin | a9075df | 2019-06-27 15:41:57 +0100 | [diff] [blame] | 1435 | IConnectableLayer* Network::AddResizeLayer(const ResizeDescriptor& |
| 1436 | resizeDescriptor, const char* name) |
| 1437 | { |
Aron Virginas-Tar | 169d2f1 | 2019-07-01 19:01:44 +0100 | [diff] [blame] | 1438 | return m_Graph->AddLayer<ResizeLayer>(resizeDescriptor, name); |
Teresa Charlin | a9075df | 2019-06-27 15:41:57 +0100 | [diff] [blame] | 1439 | } |
| 1440 | |
Kevin May | ce5045a | 2019-10-02 14:07:47 +0100 | [diff] [blame] | 1441 | IConnectableLayer* Network::AddInstanceNormalizationLayer(const InstanceNormalizationDescriptor& desc, |
| 1442 | const char* name) |
| 1443 | { |
| 1444 | return m_Graph->AddLayer<InstanceNormalizationLayer>(desc, name); |
| 1445 | } |
| 1446 | |
Matteo Martincigh | bcd3c85 | 2018-09-28 14:14:12 +0100 | [diff] [blame] | 1447 | IConnectableLayer* Network::AddL2NormalizationLayer(const L2NormalizationDescriptor& desc, |
| 1448 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1449 | { |
Matteo Martincigh | bcd3c85 | 2018-09-28 14:14:12 +0100 | [diff] [blame] | 1450 | return m_Graph->AddLayer<L2NormalizationLayer>(desc, name); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1451 | } |
| 1452 | |
Aron Virginas-Tar | f982dea | 2019-10-11 14:07:53 +0100 | [diff] [blame] | 1453 | IConnectableLayer* Network::AddLogSoftmaxLayer(const LogSoftmaxDescriptor& desc, |
| 1454 | const char* name) |
| 1455 | { |
| 1456 | return m_Graph->AddLayer<LogSoftmaxLayer>(desc, name); |
| 1457 | } |
| 1458 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1459 | IConnectableLayer* Network::AddConstantLayer(const ConstTensor& input, const char* name) |
| 1460 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1461 | auto layer = m_Graph->AddLayer<ConstantLayer>(name); |
| 1462 | |
| 1463 | layer->m_LayerOutput = std::make_unique<ScopedCpuTensorHandle>(input); |
| 1464 | |
| 1465 | return layer; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1466 | } |
| 1467 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1468 | IConnectableLayer* Network::AddReshapeLayer(const ReshapeDescriptor& reshapeDescriptor, |
| 1469 | const char* name) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1470 | { |
| 1471 | return m_Graph->AddLayer<ReshapeLayer>(reshapeDescriptor, name); |
| 1472 | } |
| 1473 | |
Nattapat Chaimanowong | 207ef9a | 2018-11-02 10:57:25 +0000 | [diff] [blame] | 1474 | IConnectableLayer* Network::AddSpaceToBatchNdLayer(const SpaceToBatchNdDescriptor& spaceToBatchNdDescriptor, |
| 1475 | const char* name) |
| 1476 | { |
| 1477 | return m_Graph->AddLayer<SpaceToBatchNdLayer>(spaceToBatchNdDescriptor, name); |
| 1478 | } |
| 1479 | |
Aron Virginas-Tar | 972af15 | 2019-06-11 14:14:03 +0100 | [diff] [blame] | 1480 | IConnectableLayer* Network::AddSpaceToDepthLayer(const SpaceToDepthDescriptor& spaceToDepthDescriptor, |
| 1481 | const char* name) |
| 1482 | { |
| 1483 | return m_Graph->AddLayer<SpaceToDepthLayer>(spaceToDepthDescriptor, name); |
| 1484 | } |
| 1485 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1486 | IConnectableLayer* Network::AddFloorLayer(const char* name) |
| 1487 | { |
| 1488 | return m_Graph->AddLayer<FloorLayer>(name); |
| 1489 | } |
| 1490 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1491 | IConnectableLayer* Network::AddLstmLayer(const LstmDescriptor& descriptor, |
| 1492 | const LstmInputParams& params, |
| 1493 | const char* name) |
| 1494 | { |
| 1495 | const auto layer = m_Graph->AddLayer<LstmLayer>(descriptor, name); |
| 1496 | |
| 1497 | //Lstm Basic Parameters |
| 1498 | layer->m_BasicParameters.m_InputToForgetWeights = |
| 1499 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToForgetWeights)); |
| 1500 | layer->m_BasicParameters.m_InputToCellWeights = |
| 1501 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToCellWeights)); |
| 1502 | layer->m_BasicParameters.m_InputToOutputWeights = |
| 1503 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToOutputWeights)); |
| 1504 | layer->m_BasicParameters.m_RecurrentToForgetWeights = |
| 1505 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToForgetWeights)); |
| 1506 | layer->m_BasicParameters.m_RecurrentToCellWeights = |
| 1507 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToCellWeights)); |
| 1508 | layer->m_BasicParameters.m_RecurrentToOutputWeights = |
| 1509 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToOutputWeights)); |
| 1510 | layer->m_BasicParameters.m_ForgetGateBias = |
| 1511 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ForgetGateBias)); |
| 1512 | layer->m_BasicParameters.m_CellBias = |
| 1513 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellBias)); |
| 1514 | layer->m_BasicParameters.m_OutputGateBias = |
| 1515 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_OutputGateBias)); |
| 1516 | |
| 1517 | //Lstm Cifg parameters |
| 1518 | if(!descriptor.m_CifgEnabled) |
| 1519 | { |
| 1520 | if(params.m_InputToInputWeights == nullptr) |
| 1521 | { |
| 1522 | throw InvalidArgumentException("AddLstmLayer: Input To Input Weights cannot be NULL"); |
| 1523 | } |
| 1524 | if(params.m_RecurrentToInputWeights == nullptr) |
| 1525 | { |
| 1526 | throw InvalidArgumentException( |
| 1527 | "AddLstmLayer: Recurrent To Input Weights cannot be NULL"); |
| 1528 | } |
| 1529 | if(params.m_InputGateBias == nullptr) |
| 1530 | { |
| 1531 | throw InvalidArgumentException("AddLstmLayer: Input Gate Bias cannot be NULL"); |
| 1532 | } |
| 1533 | layer->m_CifgParameters.m_InputToInputWeights = |
| 1534 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToInputWeights)); |
| 1535 | layer->m_CifgParameters.m_RecurrentToInputWeights = |
| 1536 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToInputWeights)); |
| 1537 | // In the VTS tests, cell-to-input weights may be null, even if the other CIFG params are not. |
| 1538 | if(params.m_CellToInputWeights != nullptr) |
| 1539 | { |
| 1540 | layer->m_CifgParameters.m_CellToInputWeights = |
| 1541 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellToInputWeights)); |
| 1542 | } |
| 1543 | layer->m_CifgParameters.m_InputGateBias = |
| 1544 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputGateBias)); |
| 1545 | } |
| 1546 | |
| 1547 | //Lstm projection parameters |
| 1548 | if(descriptor.m_ProjectionEnabled) |
| 1549 | { |
| 1550 | if(params.m_ProjectionWeights == nullptr) |
| 1551 | { |
| 1552 | throw InvalidArgumentException("AddLstmLayer: Projection Weights cannot be NULL"); |
| 1553 | } |
| 1554 | layer->m_ProjectionParameters.m_ProjectionWeights = |
| 1555 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ProjectionWeights)); |
| 1556 | if(params.m_ProjectionBias != nullptr) |
| 1557 | { |
| 1558 | layer->m_ProjectionParameters.m_ProjectionBias = |
| 1559 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ProjectionBias)); |
| 1560 | } |
| 1561 | } |
| 1562 | |
| 1563 | //Lstm Peephole params |
| 1564 | if(descriptor.m_PeepholeEnabled) |
| 1565 | { |
| 1566 | if(params.m_CellToForgetWeights == nullptr) |
| 1567 | { |
| 1568 | throw InvalidArgumentException("AddLstmLayer: Cell To Forget Weights cannot be NULL"); |
| 1569 | } |
| 1570 | if(params.m_CellToOutputWeights == nullptr) |
| 1571 | { |
| 1572 | throw InvalidArgumentException("AddLstmLayer: Cell To Output Weights cannot be NULL"); |
| 1573 | } |
| 1574 | layer->m_PeepholeParameters.m_CellToForgetWeights = |
| 1575 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellToForgetWeights)); |
| 1576 | layer->m_PeepholeParameters.m_CellToOutputWeights = |
| 1577 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellToOutputWeights)); |
| 1578 | } |
Jan Eilers | f8c6297 | 2019-07-17 11:07:49 +0100 | [diff] [blame] | 1579 | |
| 1580 | //Lstm Layer Normalization params |
| 1581 | if(descriptor.m_LayerNormEnabled) |
| 1582 | { |
| 1583 | if(!descriptor.m_CifgEnabled) |
| 1584 | { |
| 1585 | if(params.m_InputLayerNormWeights == nullptr) |
| 1586 | { |
| 1587 | throw InvalidArgumentException("AddLstmLayer: Input layer normalization weights cannot be NULL"); |
| 1588 | } |
| 1589 | layer->m_LayerNormParameters.m_InputLayerNormWeights = |
| 1590 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputLayerNormWeights)); |
| 1591 | } |
| 1592 | |
| 1593 | if(params.m_ForgetLayerNormWeights == nullptr) |
| 1594 | { |
| 1595 | throw InvalidArgumentException("AddLstmLayer: Forget layer normalization weights cannot be NULL"); |
| 1596 | } |
| 1597 | if(params.m_CellLayerNormWeights == nullptr) |
| 1598 | { |
| 1599 | throw InvalidArgumentException("AddLstmLayer: Cell layer normalization weights cannot be NULL"); |
| 1600 | } |
| 1601 | if(params.m_OutputLayerNormWeights == nullptr) |
| 1602 | { |
| 1603 | throw InvalidArgumentException("AddLstmLayer: Output layer normalization weights cannot be NULL"); |
| 1604 | } |
| 1605 | layer->m_LayerNormParameters.m_ForgetLayerNormWeights = |
| 1606 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ForgetLayerNormWeights)); |
| 1607 | layer->m_LayerNormParameters.m_CellLayerNormWeights = |
| 1608 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellLayerNormWeights)); |
| 1609 | layer->m_LayerNormParameters.m_OutputLayerNormWeights = |
| 1610 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_OutputLayerNormWeights)); |
| 1611 | } |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1612 | return layer; |
| 1613 | } |
| 1614 | |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1615 | IConnectableLayer* Network::AddDivisionLayer(const char* name) |
| 1616 | { |
| 1617 | return m_Graph->AddLayer<DivisionLayer>(name); |
| 1618 | } |
| 1619 | |
David Beck | 1952622 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1620 | IConnectableLayer* Network::AddSubtractionLayer(const char* name) |
| 1621 | { |
| 1622 | return m_Graph->AddLayer<SubtractionLayer>(name); |
| 1623 | } |
| 1624 | |
narpra01 | 32b9046 | 2018-09-13 11:07:48 +0100 | [diff] [blame] | 1625 | IConnectableLayer* Network::AddMeanLayer(const MeanDescriptor& meanDescriptor, const char* name) |
| 1626 | { |
| 1627 | return m_Graph->AddLayer<MeanLayer>(meanDescriptor,name); |
| 1628 | } |
| 1629 | |
Mohamed Nour Abouelseoud | 5662c20 | 2018-09-24 13:30:09 +0100 | [diff] [blame] | 1630 | IConnectableLayer* Network::AddPadLayer(const PadDescriptor& padDescriptor, const char* name) |
| 1631 | { |
| 1632 | return m_Graph->AddLayer<PadLayer>(padDescriptor,name); |
| 1633 | } |
| 1634 | |
Derek Lamberti | a9cca6a | 2019-03-25 15:41:58 +0000 | [diff] [blame] | 1635 | IConnectableLayer *Network::AddQuantizeLayer(const char *name) |
| 1636 | { |
| 1637 | return m_Graph->AddLayer<QuantizeLayer>(name); |
| 1638 | } |
| 1639 | |
Nattapat Chaimanowong | e4294fd | 2019-03-28 09:56:53 +0000 | [diff] [blame] | 1640 | IConnectableLayer* Network::AddDequantizeLayer(const char* name) |
| 1641 | { |
| 1642 | return m_Graph->AddLayer<DequantizeLayer>(name); |
| 1643 | } |
| 1644 | |
Conor Kennedy | 430b5d8 | 2018-11-14 15:28:28 +0000 | [diff] [blame] | 1645 | IConnectableLayer* Network::AddStridedSliceLayer(const StridedSliceDescriptor& stridedSliceDescriptor, |
| 1646 | const char* name) |
| 1647 | { |
| 1648 | return m_Graph->AddLayer<StridedSliceLayer>(stridedSliceDescriptor, name); |
| 1649 | } |
| 1650 | |
Matteo Martincigh | 59a950c | 2018-12-13 12:48:25 +0000 | [diff] [blame] | 1651 | IConnectableLayer* Network::AddGreaterLayer(const char* name) |
| 1652 | { |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 1653 | return AddComparisonLayer(ComparisonDescriptor(ComparisonOperation::Greater), name); |
Matteo Martincigh | 59a950c | 2018-12-13 12:48:25 +0000 | [diff] [blame] | 1654 | } |
| 1655 | |
FrancisMurtagh | 2099595 | 2018-12-17 12:11:36 +0000 | [diff] [blame] | 1656 | IConnectableLayer* Network::AddEqualLayer(const char* name) |
| 1657 | { |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 1658 | return AddComparisonLayer(ComparisonDescriptor(ComparisonOperation::Equal), name); |
FrancisMurtagh | 2099595 | 2018-12-17 12:11:36 +0000 | [diff] [blame] | 1659 | } |
| 1660 | |
Mohamed Nour Abouelseoud | a1d3c6a | 2018-12-27 12:39:16 +0000 | [diff] [blame] | 1661 | IConnectableLayer* Network::AddRsqrtLayer(const char * name) |
| 1662 | { |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 1663 | return AddElementwiseUnaryLayer(ElementwiseUnaryDescriptor(UnaryOperation::Rsqrt), name); |
Mohamed Nour Abouelseoud | a1d3c6a | 2018-12-27 12:39:16 +0000 | [diff] [blame] | 1664 | } |
| 1665 | |
narpra01 | b89b05f | 2019-01-16 09:53:09 +0000 | [diff] [blame] | 1666 | IConnectableLayer* Network::AddGatherLayer(const char* name) |
| 1667 | { |
| 1668 | return m_Graph->AddLayer<GatherLayer>(name); |
| 1669 | } |
| 1670 | |
Nattapat Chaimanowong | 1f88630 | 2019-04-05 13:37:19 +0100 | [diff] [blame] | 1671 | IConnectableLayer* Network::AddMergeLayer(const char* name) |
| 1672 | { |
| 1673 | return m_Graph->AddLayer<MergeLayer>(name); |
| 1674 | } |
| 1675 | |
Sadik Armagan | eff363d | 2019-04-05 15:25:46 +0100 | [diff] [blame] | 1676 | IConnectableLayer* Network::AddSwitchLayer(const char* name) |
| 1677 | { |
| 1678 | return m_Graph->AddLayer<SwitchLayer>(name); |
| 1679 | } |
| 1680 | |
Matteo Martincigh | 0e406ee | 2019-06-12 15:42:18 +0100 | [diff] [blame] | 1681 | IConnectableLayer* Network::AddPreluLayer(const char* name) |
| 1682 | { |
| 1683 | return m_Graph->AddLayer<PreluLayer>(name); |
| 1684 | } |
| 1685 | |
Aron Virginas-Tar | 639fb04 | 2019-06-20 14:28:19 +0100 | [diff] [blame] | 1686 | IConnectableLayer* Network::AddTransposeConvolution2dLayer(const TransposeConvolution2dDescriptor& descriptor, |
| 1687 | const ConstTensor& weights, |
| 1688 | const Optional<ConstTensor>& biases, |
| 1689 | const char* name) |
| 1690 | { |
| 1691 | if (descriptor.m_BiasEnabled && !biases.has_value()) |
| 1692 | { |
| 1693 | throw InvalidArgumentException("AddTransposeConvolution2dLayer: Biases cannot be empty"); |
| 1694 | } |
| 1695 | |
| 1696 | const auto layer = m_Graph->AddLayer<TransposeConvolution2dLayer>(descriptor, name); |
| 1697 | |
| 1698 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 1699 | |
| 1700 | if (descriptor.m_BiasEnabled) |
| 1701 | { |
| 1702 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(biases.value()); |
| 1703 | } |
| 1704 | |
| 1705 | return layer; |
| 1706 | } |
| 1707 | |
Mike Kelly | c9ea45a | 2020-02-28 18:11:58 +0000 | [diff] [blame] | 1708 | IConnectableLayer* Network::AddTransposeLayer(const TransposeDescriptor& transposeDescriptor, |
| 1709 | const char* name) |
| 1710 | { |
| 1711 | return m_Graph->AddLayer<TransposeLayer>(transposeDescriptor, name); |
| 1712 | } |
| 1713 | |
Matthew Jackson | 2b8c1da | 2019-07-04 14:59:16 +0100 | [diff] [blame] | 1714 | IConnectableLayer* Network::AddStackLayer(const StackDescriptor& stackDescriptor, |
| 1715 | const char* name) |
| 1716 | { |
| 1717 | return m_Graph->AddLayer<StackLayer>(stackDescriptor, name); |
| 1718 | } |
| 1719 | |
Derek Lamberti | 013c390 | 2019-10-21 10:46:16 +0100 | [diff] [blame] | 1720 | |
| 1721 | IConnectableLayer* Network::AddStandInLayer(const StandInDescriptor& desc, |
| 1722 | const char* name) |
| 1723 | { |
| 1724 | return m_Graph->AddLayer<StandInLayer>(desc, name); |
| 1725 | } |
| 1726 | |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1727 | IConnectableLayer* Network::AddQuantizedLstmLayer(const QuantizedLstmInputParams& params, |
| 1728 | const char* name) |
| 1729 | { |
| 1730 | const auto layer = m_Graph->AddLayer<QuantizedLstmLayer>(name); |
| 1731 | |
| 1732 | // InputToX weights |
| 1733 | layer->m_QuantizedLstmParameters.m_InputToInputWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1734 | std::make_unique<ScopedCpuTensorHandle>(params.GetInputToInputWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1735 | layer->m_QuantizedLstmParameters.m_InputToForgetWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1736 | std::make_unique<ScopedCpuTensorHandle>(params.GetInputToForgetWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1737 | layer->m_QuantizedLstmParameters.m_InputToCellWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1738 | std::make_unique<ScopedCpuTensorHandle>(params.GetInputToCellWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1739 | layer->m_QuantizedLstmParameters.m_InputToOutputWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1740 | std::make_unique<ScopedCpuTensorHandle>(params.GetInputToOutputWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1741 | |
| 1742 | // RecurrentToX weights |
| 1743 | layer->m_QuantizedLstmParameters.m_RecurrentToInputWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1744 | std::make_unique<ScopedCpuTensorHandle>(params.GetRecurrentToInputWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1745 | layer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1746 | std::make_unique<ScopedCpuTensorHandle>(params.GetRecurrentToForgetWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1747 | layer->m_QuantizedLstmParameters.m_RecurrentToCellWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1748 | std::make_unique<ScopedCpuTensorHandle>(params.GetRecurrentToCellWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1749 | layer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1750 | std::make_unique<ScopedCpuTensorHandle>(params.GetRecurrentToOutputWeights()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1751 | |
| 1752 | // Bias |
| 1753 | layer->m_QuantizedLstmParameters.m_InputGateBias = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1754 | std::make_unique<ScopedCpuTensorHandle>(params.GetInputGateBias()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1755 | layer->m_QuantizedLstmParameters.m_ForgetGateBias = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1756 | std::make_unique<ScopedCpuTensorHandle>(params.GetForgetGateBias()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1757 | layer->m_QuantizedLstmParameters.m_CellBias = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1758 | std::make_unique<ScopedCpuTensorHandle>(params.GetCellBias()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1759 | layer->m_QuantizedLstmParameters.m_OutputGateBias = |
Francis Murtagh | bb590b4 | 2019-08-14 09:51:36 +0100 | [diff] [blame] | 1760 | std::make_unique<ScopedCpuTensorHandle>(params.GetOutputGateBias()); |
James Conroy | ee18dc8 | 2019-07-17 11:27:46 +0100 | [diff] [blame] | 1761 | |
| 1762 | return layer; |
| 1763 | } |
| 1764 | |
James Conroy | 586a9aa | 2020-03-20 08:49:33 +0000 | [diff] [blame] | 1765 | IConnectableLayer* Network::AddQLstmLayer(const QLstmDescriptor& descriptor, |
| 1766 | const LstmInputParams& params, |
| 1767 | const char* name) |
| 1768 | { |
| 1769 | const auto layer = m_Graph->AddLayer<QLstmLayer>(descriptor, name); |
| 1770 | |
| 1771 | // QLstm Basic Parameters |
| 1772 | layer->m_BasicParameters.m_InputToForgetWeights = |
| 1773 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToForgetWeights)); |
| 1774 | layer->m_BasicParameters.m_InputToCellWeights = |
| 1775 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToCellWeights)); |
| 1776 | layer->m_BasicParameters.m_InputToOutputWeights = |
| 1777 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToOutputWeights)); |
| 1778 | layer->m_BasicParameters.m_RecurrentToForgetWeights = |
| 1779 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToForgetWeights)); |
| 1780 | layer->m_BasicParameters.m_RecurrentToCellWeights = |
| 1781 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToCellWeights)); |
| 1782 | layer->m_BasicParameters.m_RecurrentToOutputWeights = |
| 1783 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToOutputWeights)); |
| 1784 | layer->m_BasicParameters.m_ForgetGateBias = |
| 1785 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ForgetGateBias)); |
| 1786 | layer->m_BasicParameters.m_CellBias = |
| 1787 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellBias)); |
| 1788 | layer->m_BasicParameters.m_OutputGateBias = |
| 1789 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_OutputGateBias)); |
| 1790 | |
| 1791 | // QLstm Cifg parameters |
| 1792 | if(!descriptor.m_CifgEnabled) |
| 1793 | { |
| 1794 | if(params.m_InputToInputWeights == nullptr) |
| 1795 | { |
| 1796 | throw InvalidArgumentException("AddQLstmLayer: Input To Input Weights cannot be NULL"); |
| 1797 | } |
| 1798 | |
| 1799 | if(params.m_RecurrentToInputWeights == nullptr) |
| 1800 | { |
| 1801 | throw InvalidArgumentException( |
| 1802 | "AddQLstmLayer: Recurrent To Input Weights cannot be NULL"); |
| 1803 | } |
| 1804 | |
| 1805 | if(params.m_InputGateBias == nullptr) |
| 1806 | { |
| 1807 | throw InvalidArgumentException("AddQLstmLayer: Input Gate Bias cannot be NULL"); |
| 1808 | } |
| 1809 | |
| 1810 | layer->m_CifgParameters.m_InputToInputWeights = |
| 1811 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputToInputWeights)); |
| 1812 | layer->m_CifgParameters.m_RecurrentToInputWeights = |
| 1813 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_RecurrentToInputWeights)); |
| 1814 | layer->m_CifgParameters.m_InputGateBias = |
| 1815 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputGateBias)); |
| 1816 | } |
| 1817 | |
| 1818 | // QLstm Projection parameters |
| 1819 | if(descriptor.m_ProjectionEnabled) |
| 1820 | { |
| 1821 | if(params.m_ProjectionWeights == nullptr) |
| 1822 | { |
| 1823 | throw InvalidArgumentException("AddQLstmLayer: Projection Weights cannot be NULL"); |
| 1824 | } |
| 1825 | |
| 1826 | if(params.m_ProjectionBias == nullptr) |
| 1827 | { |
| 1828 | throw InvalidArgumentException("AddQLstmLayer: Projection Biases cannot be NULL"); |
| 1829 | } |
| 1830 | |
| 1831 | layer->m_ProjectionParameters.m_ProjectionWeights = |
| 1832 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ProjectionWeights)); |
| 1833 | layer->m_ProjectionParameters.m_ProjectionBias = |
| 1834 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ProjectionBias)); |
| 1835 | } |
| 1836 | |
| 1837 | // QLstm Peephole params |
| 1838 | if(descriptor.m_PeepholeEnabled) |
| 1839 | { |
| 1840 | if(params.m_CellToForgetWeights == nullptr) |
| 1841 | { |
| 1842 | throw InvalidArgumentException("AddQLstmLayer: Cell To Forget Weights cannot be NULL"); |
| 1843 | } |
| 1844 | |
| 1845 | if(params.m_CellToOutputWeights == nullptr) |
| 1846 | { |
| 1847 | throw InvalidArgumentException("AddQLstmLayer: Cell To Output Weights cannot be NULL"); |
| 1848 | } |
| 1849 | |
| 1850 | if(!descriptor.m_CifgEnabled) |
| 1851 | { |
| 1852 | if(params.m_CellToInputWeights == nullptr) |
| 1853 | { |
| 1854 | throw InvalidArgumentException("AddQLstmLayer: Cell To Input Weights cannot be NULL"); |
| 1855 | } |
| 1856 | |
| 1857 | layer->m_PeepholeParameters.m_CellToInputWeights = |
| 1858 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellToInputWeights)); |
| 1859 | } |
| 1860 | |
| 1861 | layer->m_PeepholeParameters.m_CellToForgetWeights = |
| 1862 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellToForgetWeights)); |
| 1863 | layer->m_PeepholeParameters.m_CellToOutputWeights = |
| 1864 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellToOutputWeights)); |
| 1865 | } |
| 1866 | |
| 1867 | // QLstm Layer Normalization params |
| 1868 | if(descriptor.m_LayerNormEnabled) |
| 1869 | { |
| 1870 | if(params.m_ForgetLayerNormWeights == nullptr) |
| 1871 | { |
| 1872 | throw InvalidArgumentException("AddQLstmLayer: Forget layer normalization weights cannot be NULL"); |
| 1873 | } |
| 1874 | |
| 1875 | if(params.m_CellLayerNormWeights == nullptr) |
| 1876 | { |
| 1877 | throw InvalidArgumentException("AddQLstmLayer: Cell layer normalization weights cannot be NULL"); |
| 1878 | } |
| 1879 | |
| 1880 | if(params.m_OutputLayerNormWeights == nullptr) |
| 1881 | { |
| 1882 | throw InvalidArgumentException("AddQLstmLayer: Output layer normalization weights cannot be NULL"); |
| 1883 | } |
| 1884 | |
| 1885 | if(!descriptor.m_CifgEnabled) |
| 1886 | { |
| 1887 | if(params.m_InputLayerNormWeights == nullptr) |
| 1888 | { |
| 1889 | throw InvalidArgumentException("AddQLstmLayer: Input layer normalization weights cannot be NULL"); |
| 1890 | } |
| 1891 | |
| 1892 | layer->m_LayerNormParameters.m_InputLayerNormWeights = |
| 1893 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_InputLayerNormWeights)); |
| 1894 | } |
| 1895 | |
| 1896 | layer->m_LayerNormParameters.m_ForgetLayerNormWeights = |
| 1897 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_ForgetLayerNormWeights)); |
| 1898 | layer->m_LayerNormParameters.m_CellLayerNormWeights = |
| 1899 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_CellLayerNormWeights)); |
| 1900 | layer->m_LayerNormParameters.m_OutputLayerNormWeights = |
| 1901 | std::make_unique<ScopedCpuTensorHandle>(*(params.m_OutputLayerNormWeights)); |
| 1902 | } |
| 1903 | return layer; |
| 1904 | } |
| 1905 | |
Mike Kelly | 8c1701a | 2019-02-11 17:01:27 +0000 | [diff] [blame] | 1906 | void Network::Accept(ILayerVisitor& visitor) const |
| 1907 | { |
| 1908 | for (auto layer : GetGraph()) |
| 1909 | { |
| 1910 | layer->Accept(visitor); |
| 1911 | }; |
| 1912 | } |
| 1913 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1914 | OptimizedNetwork::OptimizedNetwork(std::unique_ptr<Graph> graph) |
Sadik Armagan | 3184c90 | 2020-03-18 10:57:30 +0000 | [diff] [blame] | 1915 | : m_Graph(std::move(graph)), m_Guid(profiling::ProfilingService::GetNextGuid()) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1916 | { |
| 1917 | } |
| 1918 | |
| 1919 | OptimizedNetwork::~OptimizedNetwork() |
| 1920 | { |
| 1921 | } |
| 1922 | |
| 1923 | } // namespace armnn |