Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | |
| 7 | #include "ArmNNExecutor.hpp" |
| 8 | #include "NetworkExecutionUtils/NetworkExecutionUtils.hpp" |
| 9 | |
| 10 | #include <armnn/IAsyncExecutionCallback.hpp> |
| 11 | #include <AsyncExecutionCallback.hpp> |
| 12 | |
| 13 | |
| 14 | using namespace armnn; |
| 15 | using namespace std::chrono; |
| 16 | |
| 17 | ArmNNExecutor::ArmNNExecutor(const ExecuteNetworkParams& params, armnn::IRuntime::CreationOptions runtimeOptions) |
| 18 | : m_Params(params) |
| 19 | { |
| 20 | runtimeOptions.m_EnableGpuProfiling = params.m_EnableProfiling; |
| 21 | runtimeOptions.m_DynamicBackendsPath = params.m_DynamicBackendsPath; |
| 22 | m_Runtime = armnn::IRuntime::Create(runtimeOptions); |
| 23 | |
| 24 | auto parser = CreateParser(); |
| 25 | auto network = parser->CreateNetwork(m_Params); |
| 26 | auto optNet = OptimizeNetwork(network.get()); |
| 27 | |
| 28 | m_IOInfo = GetIOInfo(optNet.get()); |
| 29 | SetupInputsAndOutputs(); |
| 30 | |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 31 | armnn::ProfilingDetailsMethod profilingDetailsMethod = ProfilingDetailsMethod::Undefined; |
| 32 | if (params.m_OutputDetailsOnlyToStdOut) |
| 33 | { |
| 34 | profilingDetailsMethod = armnn::ProfilingDetailsMethod::DetailsOnly; |
| 35 | } |
| 36 | else if (params.m_OutputDetailsToStdOut) |
| 37 | { |
| 38 | profilingDetailsMethod = armnn::ProfilingDetailsMethod::DetailsWithEvents; |
| 39 | } |
| 40 | |
| 41 | INetworkProperties networkProperties{m_Params.m_Concurrent, |
| 42 | MemorySource::Undefined, |
| 43 | MemorySource::Undefined, |
| 44 | params.m_EnableProfiling, |
| 45 | profilingDetailsMethod}; |
| 46 | |
Colm Donelan | 7804481 | 2022-09-27 16:46:09 +0100 | [diff] [blame] | 47 | std::string errorMsg; |
| 48 | Status status = m_Runtime->LoadNetwork(m_NetworkId, std::move(optNet), errorMsg, networkProperties); |
| 49 | if (status != Status::Success) |
| 50 | { |
| 51 | std::string message("Failed to create Arm NN Executor: "); |
| 52 | message.append(errorMsg); |
| 53 | // Throwing an exception at this point in the constructor causes lots of problems. We'll instead mark this |
| 54 | // executor as not constructed. |
| 55 | ARMNN_LOG(fatal) << message; |
| 56 | m_constructionFailed = true; |
| 57 | return; |
| 58 | } |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 59 | |
| 60 | if (m_Params.m_Iterations > 1) |
| 61 | { |
| 62 | std::stringstream msg; |
| 63 | msg << "Network will be executed " << m_Params.m_Iterations; |
| 64 | if (m_Params.m_Concurrent) |
| 65 | { |
| 66 | msg << " times in an asynchronous manner. "; |
| 67 | } |
| 68 | else |
| 69 | { |
| 70 | msg << " times successively. "; |
| 71 | } |
| 72 | msg << "The input-tensor-data files will be reused recursively if the user didn't provide enough to " |
| 73 | "cover each execution."; |
| 74 | ARMNN_LOG(info) << msg.str(); |
| 75 | } |
| 76 | |
| 77 | if (m_Params.m_GenerateTensorData) |
| 78 | { |
| 79 | ARMNN_LOG(warning) << "The input data was generated, note that the output will not be useful"; |
| 80 | } |
| 81 | |
| 82 | if (m_Params.m_DontPrintOutputs) |
| 83 | { |
| 84 | ARMNN_LOG(info) << "Printing outputs to console is disabled."; |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | void ArmNNExecutor::ExecuteAsync() |
| 89 | { |
| 90 | std::vector<std::shared_ptr<armnn::IWorkingMemHandle>> memHandles; |
| 91 | std::unique_ptr<armnn::Threadpool> threadpool; |
| 92 | armnn::AsyncCallbackManager callbackManager; |
| 93 | std::unordered_map<armnn::InferenceId, const armnn::OutputTensors*> inferenceOutputMap; |
| 94 | |
| 95 | for (size_t i = 0; i < m_Params.m_ThreadPoolSize; ++i) |
| 96 | { |
| 97 | memHandles.emplace_back(m_Runtime->CreateWorkingMemHandle(m_NetworkId)); |
| 98 | } |
| 99 | |
| 100 | threadpool = std::make_unique<armnn::Threadpool>(m_Params.m_ThreadPoolSize, |
| 101 | m_Runtime.get(), |
| 102 | memHandles); |
| 103 | |
| 104 | ARMNN_LOG(info) << "Asynchronous Execution with Arm NN thread pool... \n"; |
| 105 | // Declare the latest and earliest inference times here to be used when calculating overall time |
| 106 | std::chrono::high_resolution_clock::time_point earliestStartTime = |
| 107 | std::chrono::high_resolution_clock::time_point::max(); |
| 108 | std::chrono::high_resolution_clock::time_point latestEndTime = |
| 109 | std::chrono::high_resolution_clock::now(); |
| 110 | |
| 111 | // For the asynchronous execution, we are adding a pool of working memory handles (1 per thread) in the |
| 112 | // LoadedNetwork with each scheduled inference having a specific priority |
| 113 | for (size_t i = 0; i < m_Params.m_Iterations; ++i) |
| 114 | { |
| 115 | std::shared_ptr<armnn::IProfiler> profiler = m_Runtime->GetProfiler(m_NetworkId); |
| 116 | |
| 117 | std::shared_ptr<armnn::AsyncExecutionCallback> cb = callbackManager.GetNewCallback(); |
| 118 | inferenceOutputMap.insert({cb->GetInferenceId(), &m_OutputTensorsVec[i]}); |
| 119 | threadpool->Schedule(m_NetworkId, |
| 120 | m_InputTensorsVec[i], |
| 121 | m_OutputTensorsVec[i], |
| 122 | armnn::QosExecPriority::Medium, |
| 123 | cb); |
| 124 | } |
| 125 | |
| 126 | // Check the results |
| 127 | for (size_t iteration = 0; iteration < m_Params.m_Iterations; ++iteration) |
| 128 | { |
| 129 | auto cb = callbackManager.GetNotifiedCallback(); |
| 130 | |
| 131 | // Get the results |
| 132 | if (earliestStartTime > cb->GetStartTime()) |
| 133 | { |
| 134 | earliestStartTime = cb->GetStartTime(); |
| 135 | } |
| 136 | if (latestEndTime < cb->GetEndTime()) |
| 137 | { |
| 138 | latestEndTime = cb->GetEndTime(); |
| 139 | } |
| 140 | |
| 141 | auto startTime = time_point_cast<std::chrono::milliseconds>(cb->GetStartTime()); |
| 142 | auto endTime = time_point_cast<std::chrono::milliseconds>(cb->GetEndTime()); |
| 143 | auto inferenceDuration = endTime - startTime; |
| 144 | CheckInferenceTimeThreshold(inferenceDuration, m_Params.m_ThresholdTime); |
| 145 | if(!m_Params.m_DontPrintOutputs) |
| 146 | { |
| 147 | const armnn::OutputTensors* out = inferenceOutputMap[cb->GetInferenceId()]; |
| 148 | PrintOutputTensors(out, iteration); |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | // Print duration difference between overallStartTime and overallEndTime |
| 153 | auto overallEndTime = time_point_cast<std::chrono::milliseconds>(latestEndTime); |
| 154 | auto overallStartTime = time_point_cast<std::chrono::milliseconds>(earliestStartTime); |
| 155 | auto totalInferenceDuration = overallEndTime - overallStartTime; |
| 156 | ARMNN_LOG(info) << "Overall Inference time: " << std::setprecision(2) |
| 157 | << std::fixed << totalInferenceDuration.count() << " ms\n"; |
| 158 | |
| 159 | } |
| 160 | |
| 161 | void ArmNNExecutor::ExecuteSync() |
| 162 | { |
| 163 | for (size_t x = 0; x < m_Params.m_Iterations; x++) |
| 164 | { |
| 165 | std::shared_ptr<armnn::IProfiler> profiler = m_Runtime->GetProfiler(m_NetworkId); |
| 166 | |
| 167 | const auto start_time = armnn::GetTimeNow(); |
| 168 | armnn::Status ret; |
| 169 | if (m_Params.m_ImportInputsIfAligned) |
| 170 | { |
| 171 | ret = m_Runtime->EnqueueWorkload(m_NetworkId, |
| 172 | m_InputTensorsVec[x], |
| 173 | m_OutputTensorsVec[x], |
| 174 | m_ImportedInputIds[x], |
| 175 | m_ImportedOutputIds[x]); |
| 176 | } |
| 177 | else |
| 178 | { |
| 179 | ret = m_Runtime->EnqueueWorkload(m_NetworkId, |
| 180 | m_InputTensorsVec[x], |
| 181 | m_OutputTensorsVec[x]); |
| 182 | } |
| 183 | |
| 184 | const auto inferenceDuration = armnn::GetTimeDuration(start_time); |
| 185 | |
| 186 | // If profiling is enabled print out the results |
Kevin May | 251fd95 | 2022-10-05 14:42:55 +0100 | [diff] [blame] | 187 | if(profiler && profiler->IsProfilingEnabled() && x == (m_Params.m_Iterations - 1)) |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 188 | { |
| 189 | profiler->Print(std::cout); |
| 190 | } |
| 191 | |
| 192 | if(ret == armnn::Status::Failure) |
| 193 | { |
| 194 | throw armnn::Exception("IRuntime::EnqueueWorkload failed"); |
| 195 | } |
| 196 | |
| 197 | if(!m_Params.m_DontPrintOutputs) |
| 198 | { |
| 199 | PrintOutputTensors(&m_OutputTensorsVec[x], x); |
| 200 | } |
| 201 | |
| 202 | // If thresholdTime == 0.0 (default), then it hasn't been supplied at command line |
| 203 | CheckInferenceTimeThreshold(inferenceDuration, m_Params.m_ThresholdTime); |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | std::vector<const void*> ArmNNExecutor::Execute() |
| 208 | { |
| 209 | if(m_Params.m_ThreadPoolSize == 0) |
| 210 | { |
| 211 | ExecuteSync(); |
| 212 | } |
| 213 | else |
| 214 | { |
| 215 | ExecuteAsync(); |
| 216 | } |
| 217 | std::vector<const void*> results; |
| 218 | for (auto& output : m_OutputStorage) |
| 219 | { |
| 220 | results.push_back(output.m_Mem); |
| 221 | } |
| 222 | |
| 223 | return results; |
| 224 | } |
| 225 | |
| 226 | void ArmNNExecutor::PrintNetworkInfo() |
| 227 | { |
| 228 | const std::vector<std::string>& inputNames = m_Params.m_InputNames.size() != 0 ? |
| 229 | m_Params.m_InputNames : |
| 230 | m_IOInfo.m_InputNames; |
| 231 | std::stringstream ss; |
| 232 | ss << "===== Network Info =====\n"; |
| 233 | ss << "Inputs in order:\n"; |
| 234 | for (const auto& inputName : inputNames) |
| 235 | { |
| 236 | const auto inputInfo = m_IOInfo.m_InputInfoMap[inputName].second; |
| 237 | ss << inputName << ", " << inputInfo.GetShape() << ", " << GetDataTypeName(inputInfo.GetDataType()); |
| 238 | if (inputInfo.IsQuantized()) |
| 239 | { |
| 240 | ss << " Quantization Offset: " << inputInfo.GetQuantizationOffset(); |
| 241 | if (inputInfo.HasMultipleQuantizationScales()) |
| 242 | { |
| 243 | ss << " Quantization scales: "; |
| 244 | for (const auto scale: inputInfo.GetQuantizationScales()) |
| 245 | { |
| 246 | ss << scale << ", "; |
| 247 | } |
| 248 | } |
| 249 | else |
| 250 | { |
| 251 | ss << " Quantization scale: " << inputInfo.GetQuantizationScale(); |
| 252 | } |
| 253 | } |
| 254 | ss << "\n"; |
| 255 | } |
| 256 | |
| 257 | ss << "Outputs in order:\n"; |
| 258 | for (const auto& outputName : m_IOInfo.m_OutputNames) |
| 259 | { |
| 260 | const auto outputInfo = m_IOInfo.m_OutputInfoMap[outputName].second; |
| 261 | ss << outputName << ", " << outputInfo.GetShape() << ", " << GetDataTypeName(outputInfo.GetDataType()); |
| 262 | if (outputInfo.IsQuantized()) |
| 263 | { |
| 264 | ss << " Quantization Offset: " << outputInfo.GetQuantizationOffset(); |
| 265 | if (outputInfo.HasMultipleQuantizationScales()) |
| 266 | { |
| 267 | ss << " Quantization scales: "; |
| 268 | for (const auto scale: outputInfo.GetQuantizationScales()) |
| 269 | { |
| 270 | ss << scale << ", "; |
| 271 | } |
| 272 | } |
| 273 | else |
| 274 | { |
| 275 | ss << " Quantization scale: " << outputInfo.GetQuantizationScale(); |
| 276 | } |
| 277 | } |
| 278 | ss << "\n"; |
| 279 | } |
| 280 | |
| 281 | std::cout << ss.str() << std::endl; |
| 282 | } |
| 283 | |
| 284 | void ArmNNExecutor::SetupInputsAndOutputs() |
| 285 | { |
| 286 | const unsigned int noOfInputs = m_IOInfo.m_InputNames.size(); |
| 287 | |
| 288 | if (m_Params.m_InputNames.size() != 0 && m_Params.m_InputNames.size() != noOfInputs) |
| 289 | { |
| 290 | LogAndThrow("Number of input names does not match number of inputs"); |
| 291 | } |
| 292 | |
| 293 | const unsigned int inputFilePaths = m_Params.m_InputTensorDataFilePaths.size(); |
| 294 | const std::vector<std::string>& inputNames = m_Params.m_InputNames.size() != 0 ? |
| 295 | m_Params.m_InputNames : |
| 296 | m_IOInfo.m_InputNames; |
| 297 | unsigned int noInputSets = 1; |
| 298 | |
| 299 | if (inputFilePaths != 0) |
| 300 | { |
| 301 | if (inputFilePaths % noOfInputs != 0) |
| 302 | { |
| 303 | LogAndThrow("Number of input files: " + std::to_string(inputFilePaths) + |
| 304 | " not compatible with number of inputs: " + std::to_string(noOfInputs)); |
| 305 | } |
| 306 | noInputSets = inputFilePaths / noOfInputs; |
| 307 | if (noInputSets != 1 && m_Params.m_ReuseBuffers) |
| 308 | { |
| 309 | LogAndThrow("Specifying multiple sets of inputs not compatible with ReuseBuffers"); |
| 310 | } |
| 311 | } |
| 312 | |
| 313 | const unsigned int noOfOutputs = m_IOInfo.m_OutputNames.size(); |
| 314 | const unsigned int outputFilePaths = m_Params.m_OutputTensorFiles.size(); |
| 315 | unsigned int noOutputSets = 1; |
| 316 | |
| 317 | if (outputFilePaths != 0) |
| 318 | { |
| 319 | if (outputFilePaths % noOfOutputs != 0) |
| 320 | { |
| 321 | LogAndThrow("Number of output files: " + std::to_string(outputFilePaths) + |
| 322 | ", not compatible with number of outputs: " + std::to_string(noOfOutputs)); |
| 323 | } |
| 324 | noOutputSets = outputFilePaths / noOfOutputs; |
| 325 | |
| 326 | if (noOutputSets != 1 && m_Params.m_ReuseBuffers) |
| 327 | { |
| 328 | LogAndThrow("Specifying multiple sets of outputs not compatible with ReuseBuffers"); |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | if (m_Params.m_ThreadPoolSize != 0) |
| 333 | { |
| 334 | // The current implementation of the Threadpool does not allow binding of outputs to a thread |
| 335 | // So to ensure no two threads write to the same output at the same time, no output can be reused |
| 336 | noOutputSets = m_Params.m_Iterations; |
| 337 | } |
| 338 | |
| 339 | if (m_Params.m_InputTensorDataFilePaths.size() > noOfInputs) |
| 340 | { |
| 341 | ARMNN_LOG(info) << "Given network has " << noOfInputs << " input/s. One input-tensor-data file is required " |
| 342 | << "for each input. The user provided " |
| 343 | << m_Params.m_InputTensorDataFilePaths.size() |
| 344 | << " input-tensor-data file/s which will be used to fill the input/s.\n"; |
| 345 | } |
| 346 | |
| 347 | unsigned int inputCount = 0; |
| 348 | for(unsigned int inputSet = 0; inputSet < noInputSets; ++inputSet) |
| 349 | { |
| 350 | armnn::InputTensors inputTensors; |
| 351 | for (const auto& inputName: inputNames) |
| 352 | { |
| 353 | armnn::BindingPointInfo bindingPointInfo; |
| 354 | try |
| 355 | { |
| 356 | bindingPointInfo = m_IOInfo.m_InputInfoMap.at(inputName); |
| 357 | } |
| 358 | catch (const std::out_of_range& e) |
| 359 | { |
| 360 | LogAndThrow("Input with inputName: " + inputName + " not found."); |
| 361 | } |
| 362 | |
| 363 | const armnn::TensorInfo& tensorInfo = bindingPointInfo.second; |
| 364 | auto newInfo = armnn::TensorInfo{tensorInfo.GetShape(), tensorInfo.GetDataType(), |
| 365 | tensorInfo.GetQuantizationScale(), |
| 366 | tensorInfo.GetQuantizationOffset(), |
| 367 | true}; |
| 368 | |
| 369 | m_InputStorage.emplace_back(IOStorage{tensorInfo.GetNumBytes()}); |
| 370 | |
| 371 | const int bindingId = bindingPointInfo.first; |
| 372 | inputTensors.emplace_back(bindingId, armnn::ConstTensor{newInfo, m_InputStorage.back().m_Mem}); |
| 373 | |
| 374 | const armnn::Optional<std::string> dataFile = m_Params.m_GenerateTensorData ? |
| 375 | armnn::EmptyOptional() : |
| 376 | armnn::MakeOptional<std::string>( |
| 377 | m_Params.m_InputTensorDataFilePaths.at(inputCount++)); |
| 378 | |
| 379 | switch (tensorInfo.GetDataType()) |
| 380 | { |
| 381 | case armnn::DataType::Float32: |
| 382 | { |
| 383 | auto typedTensor = reinterpret_cast<float*>(m_InputStorage.back().m_Mem); |
| 384 | PopulateTensorWithData<float>(typedTensor, tensorInfo.GetNumElements(), dataFile, inputName); |
| 385 | break; |
| 386 | } |
| 387 | case armnn::DataType::QSymmS16: |
| 388 | { |
| 389 | auto typedTensor = reinterpret_cast<int16_t*>(m_InputStorage.back().m_Mem); |
| 390 | PopulateTensorWithData<int16_t>(typedTensor, tensorInfo.GetNumElements(), dataFile, inputName); |
| 391 | break; |
| 392 | } |
| 393 | case armnn::DataType::QSymmS8: |
| 394 | case armnn::DataType::QAsymmS8: |
| 395 | { |
| 396 | auto typedTensor = reinterpret_cast<int8_t*>(m_InputStorage.back().m_Mem); |
| 397 | PopulateTensorWithData<int8_t>(typedTensor, tensorInfo.GetNumElements(), dataFile, inputName); |
| 398 | break; |
| 399 | } |
| 400 | case armnn::DataType::QAsymmU8: |
| 401 | { |
| 402 | auto typedTensor = reinterpret_cast<uint8_t*>(m_InputStorage.back().m_Mem); |
| 403 | PopulateTensorWithData<uint8_t>(typedTensor, tensorInfo.GetNumElements(), dataFile, inputName); |
| 404 | break; |
| 405 | } |
| 406 | case armnn::DataType::Signed32: |
| 407 | { |
| 408 | auto typedTensor = reinterpret_cast<int32_t*>(m_InputStorage.back().m_Mem); |
| 409 | PopulateTensorWithData<int32_t>(typedTensor, tensorInfo.GetNumElements(), dataFile, inputName); |
| 410 | break; |
| 411 | } |
| 412 | default: |
| 413 | { |
| 414 | LogAndThrow("Unexpected DataType"); |
| 415 | } |
| 416 | } |
| 417 | |
| 418 | if (m_Params.m_ImportInputsIfAligned) |
| 419 | { |
| 420 | m_ImportedInputIds.push_back( |
| 421 | m_Runtime->ImportInputs(m_NetworkId, m_InputTensorsVec.back(), armnn::MemorySource::Malloc)); |
| 422 | } |
| 423 | } |
| 424 | m_InputTensorsVec.emplace_back(inputTensors); |
| 425 | } |
| 426 | |
| 427 | for(unsigned int outputSet = 0; outputSet < noOutputSets; ++outputSet) |
| 428 | { |
| 429 | armnn::OutputTensors outputTensors; |
| 430 | for (const auto& output: m_IOInfo.m_OutputInfoMap) |
| 431 | { |
| 432 | const armnn::BindingPointInfo& bindingPointInfo = output.second; |
| 433 | const armnn::TensorInfo& tensorInfo = bindingPointInfo.second; |
| 434 | |
| 435 | m_OutputStorage.emplace_back(tensorInfo.GetNumBytes()); |
| 436 | outputTensors.emplace_back(bindingPointInfo.first, armnn::Tensor{tensorInfo, m_OutputStorage.back().m_Mem}); |
| 437 | } |
| 438 | m_OutputTensorsVec.emplace_back(outputTensors); |
| 439 | if (m_Params.m_ImportInputsIfAligned) |
| 440 | { |
| 441 | m_ImportedOutputIds.push_back( |
| 442 | m_Runtime->ImportOutputs(m_NetworkId, m_OutputTensorsVec.back(), armnn::MemorySource::Malloc)); |
| 443 | } |
| 444 | } |
| 445 | |
Teresa Charlin | 2050842 | 2022-10-26 14:03:08 +0100 | [diff] [blame] | 446 | // If iterations > noSets fill the remaining iterations repeating the given files |
| 447 | // If iterations < noSets just ignore the extra files |
| 448 | const unsigned int remainingInputSets = (m_Params.m_Iterations > noInputSets) |
| 449 | ? m_Params.m_Iterations - noInputSets |
| 450 | : 0; |
| 451 | for (unsigned int i = 0; i < remainingInputSets; ++i) |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 452 | { |
Teresa Charlin | 2050842 | 2022-10-26 14:03:08 +0100 | [diff] [blame] | 453 | m_InputTensorsVec.push_back(m_InputTensorsVec[i % noInputSets]); |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 454 | if (m_Params.m_ImportInputsIfAligned) |
| 455 | { |
Teresa Charlin | 2050842 | 2022-10-26 14:03:08 +0100 | [diff] [blame] | 456 | m_ImportedInputIds.push_back(m_ImportedInputIds[i % noInputSets]); |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 457 | } |
| 458 | } |
| 459 | |
Teresa Charlin | 2050842 | 2022-10-26 14:03:08 +0100 | [diff] [blame] | 460 | const unsigned int remainingOutputSets = (m_Params.m_Iterations > noOutputSets) |
| 461 | ? m_Params.m_Iterations - noOutputSets |
| 462 | : 0; |
| 463 | for (unsigned int i = 0; i < remainingOutputSets; ++i) |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 464 | { |
Teresa Charlin | 2050842 | 2022-10-26 14:03:08 +0100 | [diff] [blame] | 465 | m_OutputTensorsVec.push_back(m_OutputTensorsVec[i % noOutputSets]); |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 466 | if (m_Params.m_ImportInputsIfAligned) |
| 467 | { |
Teresa Charlin | 2050842 | 2022-10-26 14:03:08 +0100 | [diff] [blame] | 468 | m_ImportedOutputIds.push_back(m_ImportedOutputIds[i % noOutputSets]); |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 469 | } |
| 470 | } |
| 471 | } |
| 472 | |
| 473 | ArmNNExecutor::IOInfo ArmNNExecutor::GetIOInfo(armnn::IOptimizedNetwork* optNet) |
| 474 | { |
| 475 | struct IOStrategy : armnn::IStrategy |
| 476 | { |
| 477 | void ExecuteStrategy(const armnn::IConnectableLayer* layer, |
| 478 | const armnn::BaseDescriptor& descriptor, |
| 479 | const std::vector<armnn::ConstTensor>& constants, |
| 480 | const char* name, |
| 481 | const armnn::LayerBindingId id = 0) override |
| 482 | { |
| 483 | armnn::IgnoreUnused(descriptor, constants, id); |
| 484 | switch (layer->GetType()) |
| 485 | { |
| 486 | case armnn::LayerType::Input: |
| 487 | { |
| 488 | m_IOInfo.m_InputNames.emplace_back(name); |
| 489 | m_IOInfo.m_InputInfoMap[name] = {id, layer->GetOutputSlot(0).GetTensorInfo()}; |
| 490 | break; |
| 491 | } |
| 492 | case armnn::LayerType::Output: |
| 493 | { |
| 494 | m_IOInfo.m_OutputNames.emplace_back(name); |
| 495 | m_IOInfo.m_OutputInfoMap[name] = {id, layer->GetInputSlot(0).GetConnection()->GetTensorInfo()}; |
| 496 | break; |
| 497 | } |
| 498 | default: {} |
| 499 | } |
| 500 | } |
| 501 | IOInfo m_IOInfo; |
| 502 | }; |
| 503 | |
| 504 | IOStrategy ioStrategy; |
| 505 | optNet->ExecuteStrategy(ioStrategy); |
| 506 | |
| 507 | return ioStrategy.m_IOInfo; |
| 508 | } |
| 509 | |
| 510 | armnn::IOptimizedNetworkPtr ArmNNExecutor::OptimizeNetwork(armnn::INetwork* network) |
| 511 | { |
| 512 | armnn::IOptimizedNetworkPtr optNet{nullptr, [](armnn::IOptimizedNetwork*){}}; |
| 513 | |
| 514 | armnn::OptimizerOptions options; |
| 515 | options.m_ReduceFp32ToFp16 = m_Params.m_EnableFp16TurboMode; |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 516 | options.m_Debug = m_Params.m_PrintIntermediate; |
Keith Davis | 15f9c68 | 2022-10-14 15:50:33 +0100 | [diff] [blame] | 517 | options.m_DebugToFile = m_Params.m_PrintIntermediateOutputsToFile; |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 518 | options.m_shapeInferenceMethod = m_Params.m_InferOutputShape ? |
| 519 | armnn::ShapeInferenceMethod::InferAndValidate : |
| 520 | armnn::ShapeInferenceMethod::ValidateOnly; |
| 521 | options.m_ProfilingEnabled = m_Params.m_EnableProfiling; |
Jim Flynn | fcc72f5 | 2022-10-14 11:20:07 +0100 | [diff] [blame] | 522 | options.m_AllowExpandedDims = m_Params.m_AllowExpandedDims; |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 523 | |
| 524 | armnn::BackendOptions gpuAcc("GpuAcc", |
| 525 | { |
| 526 | { "FastMathEnabled", m_Params.m_EnableFastMath }, |
| 527 | { "SaveCachedNetwork", m_Params.m_SaveCachedNetwork }, |
| 528 | { "CachedNetworkFilePath", m_Params.m_CachedNetworkFilePath }, |
| 529 | { "MLGOTuningFilePath", m_Params.m_MLGOTuningFilePath } |
| 530 | }); |
| 531 | |
| 532 | armnn::BackendOptions cpuAcc("CpuAcc", |
| 533 | { |
| 534 | { "FastMathEnabled", m_Params.m_EnableFastMath }, |
| 535 | { "NumberOfThreads", m_Params.m_NumberOfThreads } |
| 536 | }); |
| 537 | options.m_ModelOptions.push_back(gpuAcc); |
| 538 | options.m_ModelOptions.push_back(cpuAcc); |
Jim Flynn | fcc72f5 | 2022-10-14 11:20:07 +0100 | [diff] [blame] | 539 | // The shapeInferenceMethod and allowExpandedDims values have to be added to the model options |
| 540 | // because these are what are passed to the OptimizeSubgraphViews method and are used to create |
| 541 | // the new optimized INetwork that method uses |
| 542 | armnn::BackendOptions allowExDimOpt("AllowExpandedDims", |
| 543 | { |
| 544 | { "AllowExpandedDims", m_Params.m_AllowExpandedDims } |
| 545 | }); |
| 546 | options.m_ModelOptions.push_back(allowExDimOpt); |
| 547 | armnn::BackendOptions shapeInferOpt("ShapeInferenceMethod", |
| 548 | { |
| 549 | { "InferAndValidate", m_Params.m_InferOutputShape } |
| 550 | }); |
| 551 | options.m_ModelOptions.push_back(shapeInferOpt); |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 552 | |
| 553 | const auto optimization_start_time = armnn::GetTimeNow(); |
| 554 | optNet = armnn::Optimize(*network, m_Params.m_ComputeDevices, m_Runtime->GetDeviceSpec(), options); |
| 555 | |
| 556 | ARMNN_LOG(info) << "Optimization time: " << std::setprecision(2) |
| 557 | << std::fixed << armnn::GetTimeDuration(optimization_start_time).count() << " ms\n"; |
| 558 | |
| 559 | if (!optNet) |
| 560 | { |
| 561 | LogAndThrow("Optimize returned nullptr"); |
| 562 | } |
| 563 | |
Teresa Charlin | 98d3fd8 | 2022-08-02 14:17:39 +0100 | [diff] [blame] | 564 | // If v,visualize-optimized-model is enabled then construct a file name for the dot file. |
| 565 | if (m_Params.m_EnableLayerDetails) |
| 566 | { |
| 567 | fs::path filename = m_Params.m_ModelPath; |
| 568 | filename.replace_extension("dot"); |
| 569 | std::fstream file(filename.c_str(), std::ios_base::out); |
| 570 | optNet->SerializeToDot(file); |
| 571 | } |
| 572 | |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 573 | return optNet; |
| 574 | } |
| 575 | |
| 576 | std::unique_ptr<ArmNNExecutor::IParser> ArmNNExecutor::CreateParser() |
| 577 | { |
Adam Jalkemo | 1e8187a | 2022-10-12 15:14:04 +0200 | [diff] [blame] | 578 | const fs::path modelFilename = m_Params.m_ModelPath; |
| 579 | const std::string modelExtension = modelFilename.extension(); |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 580 | |
Adam Jalkemo | 1e8187a | 2022-10-12 15:14:04 +0200 | [diff] [blame] | 581 | m_Params.m_IsModelBinary = modelExtension != ".json"; |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 582 | std::unique_ptr<IParser> parser = nullptr; |
| 583 | // Forward to implementation based on the parser type |
Adam Jalkemo | 1e8187a | 2022-10-12 15:14:04 +0200 | [diff] [blame] | 584 | if (modelExtension == ".armnn") |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 585 | { |
| 586 | #if defined(ARMNN_SERIALIZER) |
| 587 | parser = std::make_unique<ArmNNDeserializer>(); |
| 588 | #else |
| 589 | LogAndThrow("Not built with serialization support."); |
| 590 | #endif |
| 591 | } |
Adam Jalkemo | 1e8187a | 2022-10-12 15:14:04 +0200 | [diff] [blame] | 592 | else if (modelExtension == ".tflite") |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 593 | { |
| 594 | #if defined(ARMNN_TF_LITE_PARSER) |
| 595 | parser = std::make_unique<TfliteParser>(m_Params); |
| 596 | #else |
| 597 | LogAndThrow("Not built with Tensorflow-Lite parser support."); |
| 598 | #endif |
| 599 | } |
Adam Jalkemo | 1e8187a | 2022-10-12 15:14:04 +0200 | [diff] [blame] | 600 | else if (modelExtension == ".onnx") |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 601 | { |
| 602 | #if defined(ARMNN_ONNX_PARSER) |
| 603 | parser = std::make_unique<OnnxParser>(); |
| 604 | #else |
| 605 | LogAndThrow("Not built with Onnx parser support."); |
| 606 | #endif |
| 607 | } |
| 608 | |
| 609 | return parser; |
| 610 | } |
| 611 | |
| 612 | void ArmNNExecutor::PrintOutputTensors(const armnn::OutputTensors* outputTensors, |
| 613 | unsigned int iteration) |
| 614 | { |
| 615 | auto findOutputName = [&](const armnn::LayerBindingId id) |
| 616 | { |
| 617 | for (auto it = m_IOInfo.m_OutputInfoMap.begin(); it != m_IOInfo.m_OutputInfoMap.end(); ++it) |
| 618 | { |
| 619 | if (id == it->second.first) |
| 620 | { |
| 621 | return it->first; |
| 622 | } |
| 623 | } |
| 624 | return std::string{}; |
| 625 | }; |
| 626 | |
| 627 | unsigned int outputIndex = 0; |
| 628 | unsigned int numOutputs = outputTensors->size(); |
| 629 | for (const auto& output: *outputTensors) |
| 630 | { |
| 631 | const auto bindingName = findOutputName(output.first); |
| 632 | // We've made sure before that the number of output files either equals numOutputs, in which |
| 633 | // case we override those files when processing the results of each iteration (only the result |
| 634 | // of the last iteration will be stored), or there are enough |
| 635 | // output files for each output of each iteration. |
| 636 | size_t outputFileIndex = iteration * numOutputs + outputIndex; |
| 637 | if (!m_Params.m_OutputTensorFiles.empty()) |
| 638 | { |
| 639 | outputFileIndex = outputFileIndex % m_Params.m_OutputTensorFiles.size(); |
| 640 | ARMNN_LOG(info) << "Writing output: " << bindingName << " bindingId: '" |
| 641 | << output.first |
| 642 | << "' of iteration: " << iteration + 1 << " to file: '" |
| 643 | << m_Params.m_OutputTensorFiles[outputFileIndex] << "'"; |
| 644 | } |
| 645 | |
| 646 | const armnn::Optional<std::string> outputTensorFile = m_Params.m_OutputTensorFiles.empty() ? |
| 647 | armnn::EmptyOptional() : |
| 648 | armnn::MakeOptional<std::string>( |
| 649 | m_Params.m_OutputTensorFiles[outputFileIndex]); |
| 650 | |
| 651 | OutputWriteInfo outputWriteInfo |
| 652 | { |
| 653 | outputTensorFile, |
| 654 | bindingName, |
| 655 | output.second, |
| 656 | !m_Params.m_DontPrintOutputs |
| 657 | }; |
| 658 | |
| 659 | std::cout << bindingName << ": "; |
| 660 | std::vector<float> values; |
| 661 | switch (output.second.GetDataType()) |
| 662 | { |
| 663 | case armnn::DataType::Float32: |
| 664 | { |
| 665 | PrintTensor<float>(outputWriteInfo, "%f "); |
| 666 | break; |
| 667 | } |
| 668 | |
| 669 | case armnn::DataType::Signed32: |
| 670 | { |
| 671 | PrintTensor<int>(outputWriteInfo, "%d "); |
| 672 | break; |
| 673 | } |
| 674 | case armnn::DataType::QSymmS8: |
| 675 | case armnn::DataType::QAsymmS8: |
| 676 | { |
| 677 | PrintTensor<int8_t>(outputWriteInfo, "%d "); |
| 678 | break; |
| 679 | } |
| 680 | case armnn::DataType::QAsymmU8: |
| 681 | { |
| 682 | PrintTensor<uint8_t>(outputWriteInfo, "%d "); |
| 683 | break; |
| 684 | } |
| 685 | case armnn::DataType::Float16: |
| 686 | case armnn::DataType::QSymmS16: |
| 687 | case armnn::DataType::BFloat16: |
| 688 | case armnn::DataType::Boolean: |
| 689 | case armnn::DataType::Signed64: |
| 690 | default: |
| 691 | { |
| 692 | LogAndThrow("Unexpected DataType"); |
| 693 | } |
| 694 | } |
| 695 | std::cout << "\n"; |
Adam Jalkemo | 8f39363 | 2022-10-13 09:04:54 +0200 | [diff] [blame] | 696 | ++outputIndex; |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 697 | } |
| 698 | } |
| 699 | |
| 700 | void ArmNNExecutor::CompareAndPrintResult(std::vector<const void*> otherOutput) |
| 701 | { |
| 702 | unsigned int index = 0; |
| 703 | |
| 704 | for (const auto& outputTensors: m_OutputTensorsVec) |
| 705 | { |
| 706 | for (const auto& outputTensor: outputTensors) |
| 707 | { |
| 708 | float result = 0; |
| 709 | size_t size = outputTensor.second.GetNumBytes(); |
| 710 | |
| 711 | switch (outputTensor.second.GetDataType()) |
| 712 | { |
| 713 | case armnn::DataType::Float32: |
| 714 | { |
| 715 | result = ComputeRMSE<float>(outputTensor.second.GetMemoryArea(), otherOutput[index++], size); |
| 716 | break; |
| 717 | } |
Keith Davis | 45b82a5 | 2022-10-04 11:53:04 +0100 | [diff] [blame] | 718 | case armnn::DataType::Signed32: |
| 719 | { |
| 720 | result = ComputeRMSE<int32_t>(outputTensor.second.GetMemoryArea(), otherOutput[index++], size); |
| 721 | break; |
| 722 | } |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 723 | case armnn::DataType::QSymmS16: |
| 724 | { |
| 725 | result = ComputeRMSE<int16_t>(outputTensor.second.GetMemoryArea(), otherOutput[index++], size); |
| 726 | break; |
| 727 | } |
| 728 | case armnn::DataType::QSymmS8: |
Teresa Charlin | 98d3fd8 | 2022-08-02 14:17:39 +0100 | [diff] [blame] | 729 | case armnn::DataType::QAsymmS8: |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 730 | { |
| 731 | result = ComputeRMSE<int8_t>(outputTensor.second.GetMemoryArea(), otherOutput[index++], size); |
| 732 | break; |
| 733 | } |
| 734 | case armnn::DataType::QAsymmU8: |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 735 | { |
| 736 | result = ComputeRMSE<uint8_t>(outputTensor.second.GetMemoryArea(), otherOutput[index++], size); |
| 737 | break; |
| 738 | } |
| 739 | default: |
| 740 | { |
| 741 | LogAndThrow("Unexpected DataType"); |
| 742 | } |
| 743 | } |
| 744 | std::cout << "RMSE: of " << result << "\n"; |
| 745 | } |
| 746 | } |
| 747 | } |
| 748 | #if defined(ARMNN_SERIALIZER) |
| 749 | ArmNNExecutor::ArmNNDeserializer::ArmNNDeserializer() : m_Parser(armnnDeserializer::IDeserializer::Create()){} |
| 750 | |
| 751 | armnn::INetworkPtr ArmNNExecutor::ArmNNDeserializer::CreateNetwork(const ExecuteNetworkParams& params) |
| 752 | { |
| 753 | const std::string& modelPath = params.m_ModelPath; |
| 754 | |
| 755 | std::ifstream file(modelPath, std::ios::binary); |
| 756 | return m_Parser->CreateNetworkFromBinary(file); |
| 757 | } |
| 758 | |
| 759 | armnn::BindingPointInfo |
| 760 | ArmNNExecutor::ArmNNDeserializer::GetInputBindingPointInfo(size_t, const std::string& inputName) |
| 761 | { |
| 762 | armnnDeserializer::BindingPointInfo DeserializerBPI = m_Parser->GetNetworkInputBindingInfo(0, inputName); |
| 763 | return {DeserializerBPI.m_BindingId, DeserializerBPI.m_TensorInfo}; |
| 764 | } |
| 765 | |
| 766 | armnn::BindingPointInfo |
| 767 | ArmNNExecutor::ArmNNDeserializer::GetOutputBindingPointInfo(size_t, const std::string& outputName) |
| 768 | { |
| 769 | armnnDeserializer::BindingPointInfo DeserializerBPI = m_Parser->GetNetworkOutputBindingInfo(0, outputName); |
| 770 | return {DeserializerBPI.m_BindingId, DeserializerBPI.m_TensorInfo}; |
| 771 | } |
| 772 | #endif |
| 773 | |
| 774 | #if defined(ARMNN_TF_LITE_PARSER) |
| 775 | ArmNNExecutor::TfliteParser::TfliteParser(const ExecuteNetworkParams& params) |
| 776 | { |
| 777 | armnnTfLiteParser::ITfLiteParser::TfLiteParserOptions options; |
| 778 | options.m_StandInLayerForUnsupported = params.m_ParseUnsupported; |
| 779 | options.m_InferAndValidate = params.m_InferOutputShape; |
Jim Flynn | fcc72f5 | 2022-10-14 11:20:07 +0100 | [diff] [blame] | 780 | options.m_AllowExpandedDims = params.m_AllowExpandedDims; |
Teresa Charlin | 83b4291 | 2022-07-07 14:24:59 +0100 | [diff] [blame] | 781 | |
| 782 | m_Parser = armnnTfLiteParser::ITfLiteParser::Create(options); |
| 783 | } |
| 784 | |
| 785 | armnn::INetworkPtr ArmNNExecutor::TfliteParser::CreateNetwork(const ExecuteNetworkParams& params) |
| 786 | { |
| 787 | const std::string& modelPath = params.m_ModelPath; |
| 788 | return m_Parser->CreateNetworkFromBinaryFile(modelPath.c_str()); |
| 789 | } |
| 790 | |
| 791 | armnn::BindingPointInfo ArmNNExecutor::TfliteParser::GetInputBindingPointInfo(size_t subgraphId, |
| 792 | const std::string& inputName) |
| 793 | { |
| 794 | return m_Parser->GetNetworkInputBindingInfo(subgraphId, inputName); |
| 795 | } |
| 796 | |
| 797 | armnn::BindingPointInfo ArmNNExecutor::TfliteParser::GetOutputBindingPointInfo(size_t subgraphId, |
| 798 | const std::string& outputName) |
| 799 | { |
| 800 | return m_Parser->GetNetworkOutputBindingInfo(subgraphId, outputName); |
| 801 | } |
| 802 | #endif |
| 803 | |
| 804 | |
| 805 | #if defined(ARMNN_ONNX_PARSER) |
| 806 | ArmNNExecutor::OnnxParser::OnnxParser() : m_Parser(armnnOnnxParser::IOnnxParser::Create()){} |
| 807 | |
| 808 | armnn::INetworkPtr ArmNNExecutor::OnnxParser::CreateNetwork(const ExecuteNetworkParams& params) |
| 809 | { |
| 810 | const std::string& modelPath = params.m_ModelPath; |
| 811 | m_Parser = armnnOnnxParser::IOnnxParser::Create(); |
| 812 | std::map<std::string, armnn::TensorShape> inputShapes; |
| 813 | if(!params.m_InputTensorShapes.empty()) |
| 814 | { |
| 815 | const size_t numInputShapes = params.m_InputTensorShapes.size(); |
| 816 | const size_t numInputBindings = params.m_InputNames.size(); |
| 817 | if(numInputShapes < numInputBindings) |
| 818 | { |
| 819 | throw armnn::Exception( |
| 820 | fmt::format("Not every input has its tensor shape specified: expected={0}, got={1}", |
| 821 | numInputBindings, numInputShapes)); |
| 822 | } |
| 823 | |
| 824 | for (size_t i = 0; i < numInputShapes; i++) |
| 825 | { |
| 826 | inputShapes[params.m_InputNames[i]] = params.m_InputTensorShapes[i]; |
| 827 | } |
| 828 | |
| 829 | return params.m_IsModelBinary ? |
| 830 | m_Parser->CreateNetworkFromBinaryFile(modelPath.c_str(), inputShapes) : |
| 831 | m_Parser->CreateNetworkFromTextFile(modelPath.c_str(), inputShapes); |
| 832 | } |
| 833 | |
| 834 | // Handle text and binary input differently by calling the corresponding parser function |
| 835 | return params.m_IsModelBinary ? |
| 836 | m_Parser->CreateNetworkFromBinaryFile(params.m_ModelPath.c_str()) : |
| 837 | m_Parser->CreateNetworkFromTextFile(params.m_ModelPath.c_str()); |
| 838 | } |
| 839 | |
| 840 | armnn::BindingPointInfo ArmNNExecutor::OnnxParser::GetInputBindingPointInfo(size_t, const std::string& inputName) |
| 841 | { |
| 842 | return m_Parser->GetNetworkInputBindingInfo(inputName); |
| 843 | } |
| 844 | |
| 845 | armnn::BindingPointInfo ArmNNExecutor::OnnxParser::GetOutputBindingPointInfo(size_t, const std::string& outputName) |
| 846 | { |
| 847 | return m_Parser->GetNetworkOutputBindingInfo(outputName); |
| 848 | } |
| 849 | #endif |