From 626bd90378670eb5fd76f94526395430b752ad9e Mon Sep 17 00:00:00 2001 From: Francis Murtagh Date: Tue, 21 Jun 2022 13:16:23 +0000 Subject: Revert "Revert "IVGCVSW-6873 Import inputs but don't export outputs fails."" This reverts commit a0f8b15d4ddb5075f380003ff31b271d389d3b66. Reason for revert: Change-Id: Ibc4a77fa008643849da7330391942e4c87b941e2 --- delegate/include/Version.hpp | 2 +- delegate/src/armnn_delegate.cpp | 14 +- delegate/src/test/DelegateOptionsTest.cpp | 2 +- include/armnn/INetwork.hpp | 11 +- include/armnn/Version.hpp | 2 +- include/armnnOnnxParser/Version.hpp | 2 +- include/armnnTfLiteParser/Version.hpp | 2 +- python/pyarmnn/README.md | 14 +- .../examples/image_classification/README.md | 2 +- python/pyarmnn/examples/object_detection/README.md | 2 +- .../pyarmnn/examples/speech_recognition/README.md | 2 +- python/pyarmnn/src/pyarmnn/_version.py | 4 +- .../src/pyarmnn/swig/modules/armnn_network.i | 7 +- python/pyarmnn/test/test_modeloption.py | 9 +- python/pyarmnn/test/test_runtime.py | 4 +- python/pyarmnn/test/test_setup.py | 8 +- python/pyarmnn/test/test_version.py | 4 +- samples/ObjectDetection/Readme.md | 4 +- src/armnn/LoadedNetwork.cpp | 96 ++++++++++- src/armnn/Network.cpp | 21 ++- src/armnn/Network.hpp | 1 + src/armnn/Runtime.hpp | 4 +- src/armnn/test/RuntimeTests.cpp | 187 ++++++++++++++++++++- src/armnn/test/TensorHandleStrategyTest.cpp | 2 +- .../backendsCommon/test/CompatibilityTests.cpp | 2 +- .../backendsCommon/test/EndToEndTestImpl.hpp | 32 +++- .../backendsCommon/test/OptimizedNetworkTests.cpp | 2 +- src/backends/cl/test/ClCustomAllocatorTests.cpp | 1 + src/backends/cl/test/ClFallbackTests.cpp | 2 + src/backends/cl/test/ClImportTensorHandleTests.cpp | 7 + src/backends/cl/test/ClOptimizedNetworkTests.cpp | 2 +- src/backends/neon/test/NeonFallbackTests.cpp | 6 + .../neon/test/NeonOptimizedNetworkTests.cpp | 2 +- 33 files changed, 393 insertions(+), 69 deletions(-) diff --git a/delegate/include/Version.hpp b/delegate/include/Version.hpp index 34555b2c6f..c14857e320 100644 --- a/delegate/include/Version.hpp +++ b/delegate/include/Version.hpp @@ -14,7 +14,7 @@ namespace armnnDelegate // ArmNN Delegate version components #define DELEGATE_MAJOR_VERSION 26 -#define DELEGATE_MINOR_VERSION 0 +#define DELEGATE_MINOR_VERSION 1 #define DELEGATE_PATCH_VERSION 0 /// DELEGATE_VERSION: "X.Y.Z" diff --git a/delegate/src/armnn_delegate.cpp b/delegate/src/armnn_delegate.cpp index bb2f3c319a..1b6d68eb7a 100644 --- a/delegate/src/armnn_delegate.cpp +++ b/delegate/src/armnn_delegate.cpp @@ -394,14 +394,20 @@ ArmnnSubgraph* ArmnnSubgraph::Create(TfLiteContext* tfLiteContext, // Load graph into runtime std::string errorMessage; armnn::Status loadingStatus; - armnn::MemorySource memorySource = armnn::MemorySource::Undefined; + armnn::MemorySource inputSource = armnn::MemorySource::Undefined; + armnn::MemorySource outputSource = armnn::MemorySource::Undefined; + // There's a bit of an assumption here that the delegate will only support Malloc memory source. if (delegate->m_Options.GetOptimizerOptions().m_ImportEnabled) { - memorySource = armnn::MemorySource::Malloc; + inputSource = armnn::MemorySource::Malloc; + } + if (delegate->m_Options.GetOptimizerOptions().m_ExportEnabled) + { + outputSource = armnn::MemorySource::Malloc; } armnn::INetworkProperties networkProperties(false, - memorySource, - memorySource, + inputSource, + outputSource, delegate->m_Options.GetInternalProfilingState(), delegate->m_Options.GetInternalProfilingDetail()); loadingStatus = delegate->m_Runtime->LoadNetwork(networkId, diff --git a/delegate/src/test/DelegateOptionsTest.cpp b/delegate/src/test/DelegateOptionsTest.cpp index 126bf30a25..c9f1530968 100644 --- a/delegate/src/test/DelegateOptionsTest.cpp +++ b/delegate/src/test/DelegateOptionsTest.cpp @@ -173,7 +173,7 @@ TEST_CASE ("ArmnnDelegateModelOptions_CpuAcc_Test") }); modelOptions.push_back(cpuAcc); - armnn::OptimizerOptions optimizerOptions(false, false, false, false, modelOptions); + armnn::OptimizerOptions optimizerOptions(false, false, false, false, modelOptions, false); armnnDelegate::DelegateOptions delegateOptions(backends, optimizerOptions); DelegateOptionTest(::tflite::TensorType_FLOAT32, diff --git a/include/armnn/INetwork.hpp b/include/armnn/INetwork.hpp index 89b4776d39..475367ece5 100644 --- a/include/armnn/INetwork.hpp +++ b/include/armnn/INetwork.hpp @@ -144,10 +144,11 @@ struct OptimizerOptions , m_ImportEnabled(false) , m_ModelOptions() , m_ProfilingEnabled(false) + , m_ExportEnabled(false) {} OptimizerOptions(bool reduceFp32ToFp16, bool debug, bool reduceFp32ToBf16, bool importEnabled, - ModelOptions modelOptions = {}) + ModelOptions modelOptions = {}, bool exportEnabled = false) : m_ReduceFp32ToFp16(reduceFp32ToFp16) , m_Debug(debug) , m_ReduceFp32ToBf16(reduceFp32ToBf16) @@ -155,6 +156,7 @@ struct OptimizerOptions , m_ImportEnabled(importEnabled) , m_ModelOptions(modelOptions) , m_ProfilingEnabled(false) + , m_ExportEnabled(exportEnabled) { if (m_ReduceFp32ToFp16 && m_ReduceFp32ToBf16) { @@ -164,7 +166,7 @@ struct OptimizerOptions OptimizerOptions(bool reduceFp32ToFp16, bool debug, bool reduceFp32ToBf16 = false, ShapeInferenceMethod shapeInferenceMethod = armnn::ShapeInferenceMethod::ValidateOnly, - bool importEnabled = false, ModelOptions modelOptions = {}) + bool importEnabled = false, ModelOptions modelOptions = {}, bool exportEnabled = false) : m_ReduceFp32ToFp16(reduceFp32ToFp16) , m_Debug(debug) , m_ReduceFp32ToBf16(reduceFp32ToBf16) @@ -172,6 +174,7 @@ struct OptimizerOptions , m_ImportEnabled(importEnabled) , m_ModelOptions(modelOptions) , m_ProfilingEnabled(false) + , m_ExportEnabled(exportEnabled) { if (m_ReduceFp32ToFp16 && m_ReduceFp32ToBf16) { @@ -189,6 +192,7 @@ struct OptimizerOptions stream << "\tShapeInferenceMethod: " << (m_shapeInferenceMethod == ShapeInferenceMethod::ValidateOnly ? "ValidateOnly" : "InferAndValidate") << "\n"; stream << "\tImportEnabled: " << m_ImportEnabled << "\n"; + stream << "\tExportEnabled: " << m_ExportEnabled << "\n"; stream << "\tProfilingEnabled: " << m_ProfilingEnabled << "\n"; stream << "\tModelOptions: \n"; @@ -234,6 +238,9 @@ struct OptimizerOptions // Enable profiling dump of the optimizer phase bool m_ProfilingEnabled; + + // Enable Export + bool m_ExportEnabled; }; class IWorkloadFactory; diff --git a/include/armnn/Version.hpp b/include/armnn/Version.hpp index d41c4ec8af..7951eacf1d 100644 --- a/include/armnn/Version.hpp +++ b/include/armnn/Version.hpp @@ -10,7 +10,7 @@ #define STRINGIFY_MACRO(s) #s // ArmNN version components -#define ARMNN_MAJOR_VERSION 29 +#define ARMNN_MAJOR_VERSION 30 #define ARMNN_MINOR_VERSION 0 #define ARMNN_PATCH_VERSION 0 diff --git a/include/armnnOnnxParser/Version.hpp b/include/armnnOnnxParser/Version.hpp index ed9d8690ec..33a2846263 100644 --- a/include/armnnOnnxParser/Version.hpp +++ b/include/armnnOnnxParser/Version.hpp @@ -14,7 +14,7 @@ namespace armnnOnnxParser // OnnxParser version components #define ONNX_PARSER_MAJOR_VERSION 24 -#define ONNX_PARSER_MINOR_VERSION 4 +#define ONNX_PARSER_MINOR_VERSION 5 #define ONNX_PARSER_PATCH_VERSION 0 /// ONNX_PARSER_VERSION: "X.Y.Z" diff --git a/include/armnnTfLiteParser/Version.hpp b/include/armnnTfLiteParser/Version.hpp index eee2124678..5db527ec8c 100644 --- a/include/armnnTfLiteParser/Version.hpp +++ b/include/armnnTfLiteParser/Version.hpp @@ -14,7 +14,7 @@ namespace armnnTfLiteParser // TfLiteParser version components #define TFLITE_PARSER_MAJOR_VERSION 24 -#define TFLITE_PARSER_MINOR_VERSION 4 +#define TFLITE_PARSER_MINOR_VERSION 5 #define TFLITE_PARSER_PATCH_VERSION 0 /// TFLITE_PARSER_VERSION: "X.Y.Z" diff --git a/python/pyarmnn/README.md b/python/pyarmnn/README.md index 7dc8d8693a..6d8b42d41a 100644 --- a/python/pyarmnn/README.md +++ b/python/pyarmnn/README.md @@ -91,14 +91,14 @@ This step will put all generated files under `./src/pyarmnn/_generated` folder a ```bash $ python setup.py sdist ``` -As the result you will get `./dist/pyarmnn-29.0.0.tar.gz` file. As you can see it is platform independent. +As the result you will get `./dist/pyarmnn-30.0.0.tar.gz` file. As you can see it is platform independent. ##### 5. Build the binary package ```bash $ python setup.py bdist_wheel ``` -As the result you will get something like `./dist/pyarmnn-29.0.0-cp36-cp36m-linux_x86_64.whl` file. As you can see it +As the result you will get something like `./dist/pyarmnn-30.0.0-cp36-cp36m-linux_x86_64.whl` file. As you can see it is platform dependent. # PyArmNN installation @@ -107,8 +107,8 @@ PyArmNN can be distributed as a source package or a binary package (wheel). Binary package is platform dependent, the name of the package will indicate the platform it was built for, e.g.: -* Linux x86 64bit machine: pyarmnn-29.0.0-cp36-cp36m-*linux_x86_64*.whl -* Linux Aarch 64 bit machine: pyarmnn-29.0.0-cp36-cp36m-*linux_aarch64*.whl +* Linux x86 64bit machine: pyarmnn-30.0.0-cp36-cp36m-*linux_x86_64*.whl +* Linux Aarch 64 bit machine: pyarmnn-30.0.0-cp36-cp36m-*linux_aarch64*.whl The source package is platform independent but installation involves compilation of Arm NN python extension. You will need to have g++ compatible with C++ 14 standard and a python development library installed on the build machine. @@ -126,7 +126,7 @@ $ gcc --print-search-dirs ``` Install PyArmNN from binary by pointing to the wheel file: ```bash -$ pip install /path/to/pyarmnn-29.0.0-cp36-cp36m-linux_aarch64.whl +$ pip install /path/to/pyarmnn-30.0.0-cp36-cp36m-linux_aarch64.whl ``` ## Installing from source package @@ -143,7 +143,7 @@ $ export ARMNN_INCLUDE=/full/path/to/armnn/include:/full/path/to/armnn/profilin Install PyArmNN as follows: ```bash -$ pip install /path/to/pyarmnn-29.0.0.tar.gz +$ pip install /path/to/pyarmnn-30.0.0.tar.gz ``` If PyArmNN installation script fails to find Arm NN libraries it will raise an error like this @@ -157,7 +157,7 @@ $ pip show pyarmnn You can also verify it by running the following and getting output similar to below: ```bash $ python -c "import pyarmnn as ann;print(ann.GetVersion())" -'29.0.0' +'30.0.0' ``` # PyArmNN API overview diff --git a/python/pyarmnn/examples/image_classification/README.md b/python/pyarmnn/examples/image_classification/README.md index 7275a2523f..a360f01148 100644 --- a/python/pyarmnn/examples/image_classification/README.md +++ b/python/pyarmnn/examples/image_classification/README.md @@ -20,7 +20,7 @@ $ pip show pyarmnn You can also verify it by running the following and getting output similar to below: ```bash $ python -c "import pyarmnn as ann;print(ann.GetVersion())" -'29.0.0' +'30.0.0' ``` ##### Dependencies diff --git a/python/pyarmnn/examples/object_detection/README.md b/python/pyarmnn/examples/object_detection/README.md index 7a946ad6f5..215cf772a2 100644 --- a/python/pyarmnn/examples/object_detection/README.md +++ b/python/pyarmnn/examples/object_detection/README.md @@ -54,7 +54,7 @@ $ pip show pyarmnn You can also verify it by running the following and getting output similar to below: ```bash $ python -c "import pyarmnn as ann;print(ann.GetVersion())" -'29.0.0' +'30.0.0' ``` ##### Dependencies diff --git a/python/pyarmnn/examples/speech_recognition/README.md b/python/pyarmnn/examples/speech_recognition/README.md index 854cdaf03b..d5fee8a010 100644 --- a/python/pyarmnn/examples/speech_recognition/README.md +++ b/python/pyarmnn/examples/speech_recognition/README.md @@ -18,7 +18,7 @@ You can also verify it by running the following and getting output similar to be ```bash $ python -c "import pyarmnn as ann;print(ann.GetVersion())" -'29.0.0' +'30.0.0' ``` ### Dependencies diff --git a/python/pyarmnn/src/pyarmnn/_version.py b/python/pyarmnn/src/pyarmnn/_version.py index 7c0940e7f3..d1b1ca290c 100644 --- a/python/pyarmnn/src/pyarmnn/_version.py +++ b/python/pyarmnn/src/pyarmnn/_version.py @@ -3,7 +3,7 @@ # SPDX-License-Identifier: MIT import os -version_info = (29, 0, 0) +version_info = (30, 0, 0) __dev_version_env = os.getenv("PYARMNN_DEV_VER", "") @@ -24,7 +24,7 @@ def check_armnn_version(installed_armnn_version: str, expected_armnn_version: st """Compares expected Arm NN version and Arm NN version used to build the package. Args: - installed_armnn_version (str): Arm NN version used to generate the package (e.g. 29.0.0) + installed_armnn_version (str): Arm NN version used to generate the package (e.g. 30.0.0) expected_armnn_version (str): Expected Arm NN version Returns: diff --git a/python/pyarmnn/src/pyarmnn/swig/modules/armnn_network.i b/python/pyarmnn/src/pyarmnn/swig/modules/armnn_network.i index a2f57a3aa9..55b6795c90 100644 --- a/python/pyarmnn/src/pyarmnn/swig/modules/armnn_network.i +++ b/python/pyarmnn/src/pyarmnn/swig/modules/armnn_network.i @@ -29,7 +29,7 @@ Contains: that can not be reduced will be left in Fp32. m_ReduceFp32ToFp16 (bool): Reduces Fp32 network to Fp16 for faster processing. Layers that can not be reduced will be left in Fp32. - m_ImportEnabled (bool): Enable memory import. + m_ImportEnabled (bool): Enable memory import of inport tensors. m_shapeInferenceMethod: The ShapeInferenceMethod modifies how the output shapes are treated. When ValidateOnly is selected, the output shapes are inferred from the input parameters of the layer and any mismatch is reported. @@ -38,6 +38,7 @@ Contains: with tensors which rank or dimension sizes are not specified explicitly, however this information can be calculated from the inputs. m_ModelOptions: List of backends optimisation options. + m_ExportEnabled (bool): Enable memory export of output tensors. ") OptimizerOptions; @@ -51,7 +52,8 @@ struct OptimizerOptions bool reduceFp32ToBf16 = false, ShapeInferenceMethod shapeInferenceMethod = armnn::ShapeInferenceMethod::ValidateOnly, bool importEnabled = false, - std::vector modelOptions = {}); + std::vector modelOptions = {}, + bool exportEnabled = false); bool m_ReduceFp32ToBf16; bool m_ReduceFp32ToFp16; @@ -59,6 +61,7 @@ struct OptimizerOptions ShapeInferenceMethod m_shapeInferenceMethod; bool m_ImportEnabled; std::vector m_ModelOptions; + bool m_ExportEnabled; }; %model_options_clear; diff --git a/python/pyarmnn/test/test_modeloption.py b/python/pyarmnn/test/test_modeloption.py index c03d4a8cce..a47d2da358 100644 --- a/python/pyarmnn/test/test_modeloption.py +++ b/python/pyarmnn/test/test_modeloption.py @@ -71,7 +71,8 @@ def test_optimizer_options_with_model_opt(): False, ShapeInferenceMethod_InferAndValidate, True, - [a]) + [a], + True) mo = oo.m_ModelOptions @@ -112,7 +113,8 @@ def test_optimizer_options_fail(): False, ShapeInferenceMethod_InferAndValidate, True, - a) + a, + True) assert "Wrong number or type of arguments" in str(err.value) @@ -122,7 +124,8 @@ def test_optimizer_options_fail(): True, ShapeInferenceMethod_InferAndValidate, True, - [a]) + [a], + True) assert "BFloat16 and Float16 optimization cannot be enabled at the same time" in str(err.value) diff --git a/python/pyarmnn/test/test_runtime.py b/python/pyarmnn/test/test_runtime.py index a37772c5df..a6c4e1dcc0 100644 --- a/python/pyarmnn/test/test_runtime.py +++ b/python/pyarmnn/test/test_runtime.py @@ -156,8 +156,8 @@ def test_load_network_properties_provided(random_runtime): opt_network, _ = ann.Optimize(network, preferred_backends, runtime.GetDeviceSpec(), ann.OptimizerOptions()) - inputSource = ann.MemorySource_Malloc - outputSource = ann.MemorySource_Malloc + inputSource = ann.MemorySource_Undefined + outputSource = ann.MemorySource_Undefined properties = ann.INetworkProperties(False, inputSource, outputSource) net_id, messages = runtime.LoadNetwork(opt_network, properties) assert "" == messages diff --git a/python/pyarmnn/test/test_setup.py b/python/pyarmnn/test/test_setup.py index 4a6f930cbf..27feda2647 100644 --- a/python/pyarmnn/test/test_setup.py +++ b/python/pyarmnn/test/test_setup.py @@ -87,15 +87,15 @@ def test_gcc_serch_path(): def test_armnn_version(): - check_armnn_version('29.0.0', '29.0.0') + check_armnn_version('30.0.0', '30.0.0') def test_incorrect_armnn_version(): with pytest.raises(AssertionError) as err: - check_armnn_version('29.0.0', '29.1.0') + check_armnn_version('30.0.0', '30.1.0') - assert 'Expected ArmNN version is 29.1.0 but installed ArmNN version is 29.0.0' in str(err.value) + assert 'Expected ArmNN version is 30.1.0 but installed ArmNN version is 30.0.0' in str(err.value) def test_armnn_version_patch_does_not_matter(): - check_armnn_version('29.0.0', '29.0.1') + check_armnn_version('30.0.0', '30.0.1') diff --git a/python/pyarmnn/test/test_version.py b/python/pyarmnn/test/test_version.py index f74ae020cf..83606ab15b 100644 --- a/python/pyarmnn/test/test_version.py +++ b/python/pyarmnn/test/test_version.py @@ -18,7 +18,7 @@ def test_dev_version(): importlib.reload(v) - assert "29.0.0.dev1" == v.__version__ + assert "30.0.0.dev1" == v.__version__ del os.environ["PYARMNN_DEV_VER"] del v @@ -30,7 +30,7 @@ def test_arm_version_not_affected(): importlib.reload(v) - assert "29.0.0" == v.__arm_ml_version__ + assert "30.0.0" == v.__arm_ml_version__ del os.environ["PYARMNN_DEV_VER"] del v diff --git a/samples/ObjectDetection/Readme.md b/samples/ObjectDetection/Readme.md index 194a3e918d..bd84e26001 100644 --- a/samples/ObjectDetection/Readme.md +++ b/samples/ObjectDetection/Readme.md @@ -253,8 +253,8 @@ From the build directory, copy the following to the host platform: The full list of libs after cross-compilation to copy on your board: ``` libarmnn.so -libarmnn.so.29 -libarmnn.so.29.0 +libarmnn.so.30 +libarmnn.so.30.0 For Arm NN public C++ API mode: libarmnnTfLiteParser.so libarmnnTfLiteParser.so.24.4 diff --git a/src/armnn/LoadedNetwork.cpp b/src/armnn/LoadedNetwork.cpp index ec79d5da3e..a27add921e 100644 --- a/src/armnn/LoadedNetwork.cpp +++ b/src/armnn/LoadedNetwork.cpp @@ -84,6 +84,87 @@ void AddWorkloadStructure(std::unique_ptr& timelineUtils } // anonymous +/** + * This function performs a sanity check to ensure that the combination of input and output memory source matches the + * values for importEnabled and exportEnabled that were specified during optimization. During optimization the tensor + * handle factories are chosen based on whether import and export are enabled. If the user then specifies something + * incompatible here it can lead to problems. + * + * @param optimizedOptions + * @param networkProperties + */ +void ValidateSourcesMatchOptimizedNetwork(std::vector optimizedOptions, + const INetworkProperties& networkProperties) +{ + // Find the "Global" backend options. During the optimize phase the values of importEnabled and exportEnabled are + // added as backend options. + const vector::iterator& backendItr = + find_if(optimizedOptions.begin(), optimizedOptions.end(), [](const BackendOptions& backend) { + if (backend.GetBackendId().Get() == "Global") + { + return true; + } + else + { + return false; + } + }); + bool importEnabled = false; + bool exportEnabled = false; + if (backendItr != optimizedOptions.end()) + { + // Find the importEnabled and exportEnabled values. + for (size_t i = 0; i < backendItr->GetOptionCount(); i++) + { + const BackendOptions::BackendOption& option = backendItr->GetOption(i); + if (option.GetName() == "ImportEnabled") + { + importEnabled = option.GetValue().AsBool(); + } + if (option.GetName() == "ExportEnabled") + { + exportEnabled = option.GetValue().AsBool(); + } + } + } + + // Now that we have values for import and export compare them to the MemorySource variables. + // Any value of MemorySource that's not "Undefined" implies that we need to do an import of some kind. + if ((networkProperties.m_InputSource == MemorySource::Undefined && importEnabled) || + (networkProperties.m_InputSource != MemorySource::Undefined && !importEnabled)) + { + auto message = fmt::format("The input memory source specified, '{0}',", networkProperties.m_InputSource); + if (!importEnabled) + { + message.append(" requires that memory import be enabled. However, " + "it was disabled when this network was optimized."); + } + else + { + message.append(" requires that memory import be disabled. However, " + "it was enabled when this network was optimized."); + } + throw InvalidArgumentException(message); + } + + if ((networkProperties.m_OutputSource == MemorySource::Undefined && exportEnabled) || + (networkProperties.m_OutputSource != MemorySource::Undefined && !exportEnabled)) + { + auto message = fmt::format("The output memory source specified, '{0}',", networkProperties.m_OutputSource); + if (!exportEnabled) + { + message.append(" requires that memory export be enabled. However, " + "it was disabled when this network was optimized."); + } + else + { + message.append(" requires that memory export be disabled. However, " + "it was enabled when this network was optimized."); + } + throw InvalidArgumentException(message); + } +} // anonymous + std::unique_ptr LoadedNetwork::MakeLoadedNetwork(std::unique_ptr net, std::string& errorMessage, const INetworkProperties& networkProperties, @@ -136,6 +217,11 @@ LoadedNetwork::LoadedNetwork(std::unique_ptr net, profiler->EnableNetworkDetailsToStdOut(networkProperties.m_OutputNetworkDetailsMethod); + // We need to check that the memory sources match up with the values of import and export specified during the + // optimize phase. If they don't this will throw an exception. + ValidateSourcesMatchOptimizedNetwork(m_OptimizedNetwork.get()->pOptimizedNetworkImpl->GetModelOptions(), + m_NetworkProperties); + //First create tensor handlers, backends and workload factories. //Handlers are created before workloads are. //Because workload creation can modify some of the handlers, @@ -1439,7 +1525,7 @@ std::vector LoadedNetwork::ImportInputs(const InputTensors& inp ITensorHandle* tensorHandle = importedTensorHandlePin.m_TensorHandle.get(); - if (!CheckFlag(tensorHandle->GetImportFlags(), m_NetworkProperties.m_InputSource)) + if (!CheckFlag(tensorHandle->GetImportFlags(), forceImportMemorySource)) { throw MemoryImportException( fmt::format("ImportInputs: Memory Import failed, backend: " @@ -1451,7 +1537,7 @@ std::vector LoadedNetwork::ImportInputs(const InputTensors& inp std::make_unique(inputTensor.second.GetInfo(), inputTensor.second.GetMemoryArea()); - if (tensorHandle->Import(passThroughTensorHandle->Map(), m_NetworkProperties.m_InputSource)) + if (tensorHandle->Import(passThroughTensorHandle->Map(), forceImportMemorySource)) { importedInputs.push_back(m_CurImportedInputId++); passThroughTensorHandle->Unmap(); @@ -1564,14 +1650,14 @@ std::vector LoadedNetwork::ImportOutputs(const OutputTensors& ITensorHandle* tensorHandle = importedTensorHandlePin.m_TensorHandle.get(); - if (!CheckFlag(tensorHandle->GetImportFlags(), m_NetworkProperties.m_OutputSource)) + if (!CheckFlag(tensorHandle->GetImportFlags(), forceImportMemorySource)) { throw MemoryImportException(fmt::format("ImportInputs: Memory Import failed, backend: " "{} does not support importing from source {}" - , factoryId, m_NetworkProperties.m_OutputSource)); + , factoryId, forceImportMemorySource)); } - if (tensorHandle->Import(outputTensor.second.GetMemoryArea(), m_NetworkProperties.m_OutputSource)) + if (tensorHandle->Import(outputTensor.second.GetMemoryArea(), forceImportMemorySource)) { importedOutputs.push_back(m_CurImportedOutputId++); } diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp index f2ba94f597..9520c1399e 100644 --- a/src/armnn/Network.cpp +++ b/src/armnn/Network.cpp @@ -1362,7 +1362,7 @@ ITensorHandleFactory::FactoryId CalculateSlotOptionForOutput(BackendsMap& backen ITensorHandleFactory::FactoryId CalculateSlotOption(BackendsMap& backends, OutputSlot& outputSlot, TensorHandleFactoryRegistry& registry, - bool importEnabled) + bool exportEnabled) { // First ensure the from backends can support the TensorHandeAPI Layer& layer = outputSlot.GetOwningLayer(); @@ -1390,7 +1390,7 @@ ITensorHandleFactory::FactoryId CalculateSlotOption(BackendsMap& backends, std::map factoryScores; for (auto&& pref : srcPrefs) { - if (importEnabled) + if (exportEnabled) { ITensorHandleFactory* factory = registry.GetFactory(pref); if (outputConnection) @@ -1602,12 +1602,13 @@ OptimizationResult SelectTensorHandleStrategy(Graph& optGraph, BackendsMap& backends, TensorHandleFactoryRegistry& registry, bool importEnabled, + bool exportEnabled, Optional&> errMessages) { ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "Optimizer_SelectTensorHandleStrategy"); OptimizationResult result; - optGraph.ForEachLayer([&backends, ®istry, &result, &errMessages, importEnabled](Layer* layer) + optGraph.ForEachLayer([&backends, ®istry, &result, &errMessages, importEnabled, exportEnabled](Layer* layer) { ARMNN_ASSERT(layer); @@ -1632,7 +1633,7 @@ OptimizationResult SelectTensorHandleStrategy(Graph& optGraph, slotOption = CalculateSlotOptionForOutput(backends, outputSlot, registry); break; default: - slotOption = CalculateSlotOption(backends, outputSlot, registry, importEnabled); + slotOption = CalculateSlotOption(backends, outputSlot, registry, exportEnabled); break; } outputSlot.SetTensorHandleFactory(slotOption); @@ -1696,7 +1697,15 @@ IOptimizedNetworkPtr Optimize(const Graph& inGraph, std::unique_ptr graph = std::make_unique(inGraph); - auto optNet = IOptimizedNetworkPtr(new IOptimizedNetwork(std::move(graph), options.m_ModelOptions), + // We need to pass on the information about whether import and export is enabled to the LoadNetwork phase. + // The mechanism to do that is to add model options to the optimized network. + armnn::BackendOptions importExport("Global", + {{"ImportEnabled", options.m_ImportEnabled}, + {"ExportEnabled", options.m_ExportEnabled}}); + ModelOptions optimizedOptions(options.m_ModelOptions); + optimizedOptions.push_back(importExport); + + auto optNet = IOptimizedNetworkPtr(new IOptimizedNetwork(std::move(graph), optimizedOptions), &IOptimizedNetwork::Destroy); IOptimizedNetwork* optNetObjPtr = optNet.get(); @@ -1819,7 +1828,9 @@ IOptimizedNetworkPtr Optimize(const Graph& inGraph, backends, tensorHandleFactoryRegistry, options.m_ImportEnabled, + options.m_ExportEnabled, messages); + if (strategyResult.m_Error) { // Failed to apply the backend-specific optimizations diff --git a/src/armnn/Network.hpp b/src/armnn/Network.hpp index 6c7c2f5c7e..2d34cfc3e2 100644 --- a/src/armnn/Network.hpp +++ b/src/armnn/Network.hpp @@ -300,6 +300,7 @@ OptimizationResult SelectTensorHandleStrategy(Graph& optGraph, BackendsMap& backends, TensorHandleFactoryRegistry& registry, bool importEnabled, + bool exportEnabled, Optional&> errMessages); OptimizationResult AssignBackends(OptimizedNetworkImpl* optNetObjPtr, diff --git a/src/armnn/Runtime.hpp b/src/armnn/Runtime.hpp index 376cdbc000..f5dfadf948 100644 --- a/src/armnn/Runtime.hpp +++ b/src/armnn/Runtime.hpp @@ -56,9 +56,9 @@ public: armnn::TensorInfo GetOutputTensorInfo(NetworkId networkId, LayerBindingId layerId) const; std::vector ImportInputs(NetworkId networkId, const InputTensors& inputTensors, - MemorySource forceImportMemorySource = MemorySource::Undefined); + MemorySource forceImportMemorySource); std::vector ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors, - MemorySource forceImportMemorySource = MemorySource::Undefined); + MemorySource forceImportMemorySource); void ClearImportedInputs(NetworkId networkId, const std::vector inputIds); void ClearImportedOutputs(NetworkId networkId, const std::vector outputIds); diff --git a/src/armnn/test/RuntimeTests.cpp b/src/armnn/test/RuntimeTests.cpp index 3cbe8848df..59f65541b8 100644 --- a/src/armnn/test/RuntimeTests.cpp +++ b/src/armnn/test/RuntimeTests.cpp @@ -93,7 +93,7 @@ TEST_CASE("RuntimePreImportInputs") std::vector backends = {armnn::Compute::CpuRef}; std::string er; - armnn::INetworkProperties networkProperties(true, MemorySource::Malloc, MemorySource::Undefined); + armnn::INetworkProperties networkProperties(true, MemorySource::Undefined, MemorySource::Undefined); runtime->LoadNetwork(networkId, Optimize(*testNetwork, backends, runtime->GetDeviceSpec()), er, @@ -107,7 +107,7 @@ TEST_CASE("RuntimePreImportInputs") ConstTensor inputTensor2({{4}, armnn::DataType::Signed32, 0.0f, 0, true}, inputData2.data()); Tensor outputTensor({{4}, armnn::DataType::Signed32}, output.data()); - auto importedInputVec1 = runtime->ImportInputs(networkId, {{0, inputTensor1}}); + auto importedInputVec1 = runtime->ImportInputs(networkId, {{0, inputTensor1}}, MemorySource::Malloc); CHECK(importedInputVec1.size() == 1); CHECK(importedInputVec1[0] == 0); @@ -118,7 +118,7 @@ TEST_CASE("RuntimePreImportInputs") CHECK(val == 30); } - auto importedInputVec2 = runtime->ImportInputs(networkId, {{1, inputTensor2}}); + auto importedInputVec2 = runtime->ImportInputs(networkId, {{1, inputTensor2}}, MemorySource::Malloc); CHECK(importedInputVec2.size() == 1); CHECK(importedInputVec2[0] == 1); @@ -146,7 +146,7 @@ TEST_CASE("RuntimePreImportInputs") // Incorrect layer binding id and ImportedInputId CHECK_THROWS_AS(runtime->Execute(*memHandle.get(), {{-2, inputTensor2}}, {{2, outputTensor}}, {10});, armnn::InvalidArgumentException); - auto importedInputVec3 = runtime->ImportInputs(networkId, {{1, inputTensor2}}); + auto importedInputVec3 = runtime->ImportInputs(networkId, {{1, inputTensor2}}, MemorySource::Malloc); CHECK(importedInputVec3[0] == 2); // Too many ImportedInputIds CHECK_THROWS_AS(runtime->Execute(*memHandle.get(), {}, {{2, outputTensor}}, {0, 1, 2});, @@ -175,6 +175,7 @@ TEST_CASE("RuntimePreImportInputs") // Trying to delete unknown pre-imported tensor CHECK_THROWS_AS(runtime->ClearImportedInputs(networkId, {10});, armnn::InvalidArgumentException); } + TEST_CASE("RuntimePreImportOutputs") { armnn::IRuntime::CreationOptions options; @@ -216,7 +217,7 @@ TEST_CASE("RuntimePreImportOutputs") std::vector backends = { armnn::Compute::CpuRef }; std::string er; - armnn::INetworkProperties networkProperties(true, MemorySource::Malloc, MemorySource::Malloc); + armnn::INetworkProperties networkProperties(true, MemorySource::Undefined, MemorySource::Undefined); runtime->LoadNetwork(networkId, Optimize(*testNetwork, backends, runtime->GetDeviceSpec()), er, @@ -257,7 +258,7 @@ TEST_CASE("RuntimePreImportOutputs") runtime->Execute(*memHandle.get(),inputTensors, {output1, output2}); testOutputs(); - auto importedOutputVec = runtime->ImportOutputs(networkId, {output1, output2 }); + auto importedOutputVec = runtime->ImportOutputs(networkId, {output1, output2 }, MemorySource::Malloc); CHECK(importedOutputVec.size() == 2); CHECK(importedOutputVec[0] == 0); CHECK(importedOutputVec[1] == 1); @@ -271,7 +272,7 @@ TEST_CASE("RuntimePreImportOutputs") runtime->Execute(*memHandle.get(), inputTensors, {output2}, {}, {0}); testOutputs(); - auto importedInputVec = runtime->ImportInputs(networkId, inputTensors); + auto importedInputVec = runtime->ImportInputs(networkId, inputTensors, MemorySource::Malloc); CHECK(importedInputVec.size() == 2); CHECK(importedInputVec[0] == 0); CHECK(importedInputVec[1] == 1); @@ -1293,4 +1294,176 @@ TEST_CASE("ProfilingPostOptimisationStructureCpuRef") VerifyPostOptimisationStructureTestImpl(armnn::Compute::CpuRef); } +TEST_CASE("RuntimeOptimizeImportOff_LoadNetworkImportOn") +{ + // In this test case we'll optimize a network with both import and export disabled. Then we'll attempt to load + // that network but specify that the import memory source is Malloc. + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + armnn::NetworkId networkId = 1; + armnn::INetworkPtr testNetwork(armnn::INetwork::Create()); + + auto inputLayer1 = testNetwork->AddInputLayer(0, "input 1 layer"); + auto inputLayer2 = testNetwork->AddInputLayer(1, "input 2 layer"); + auto addLayer = testNetwork->AddAdditionLayer("add layer"); + auto outputLayer = testNetwork->AddOutputLayer(2, "output layer"); + + TensorInfo tensorInfo{{4}, armnn::DataType::Signed32}; + + inputLayer1->GetOutputSlot(0).Connect(addLayer->GetInputSlot(0)); + inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + inputLayer2->GetOutputSlot(0).Connect(addLayer->GetInputSlot(1)); + inputLayer2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + addLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + addLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + std::vector backends = {armnn::Compute::CpuRef}; + + OptimizerOptions optimizedOptions; + // Hard set import and export to off. + optimizedOptions.m_ImportEnabled = false; + optimizedOptions.m_ExportEnabled = false; + IOptimizedNetworkPtr optNet = Optimize(*testNetwork, backends, runtime->GetDeviceSpec(), optimizedOptions); + CHECK(optNet); + + std::string er; + // Load the network passing an import memory source. + armnn::INetworkProperties networkProperties1(true, MemorySource::Malloc, MemorySource::Undefined); + // There should be an InvalidArgumentException. + runtime->LoadNetwork(networkId, std::move(optNet), er, networkProperties1); + CHECK(er.find("However, it was disabled when this network was optimized") != -1); +} + +TEST_CASE("RuntimeOptimizeExportOff_LoadNetworkExportOn") +{ + // In this test case we'll optimize a network with both import and export disabled. Then we'll attempt to load + // that network but specify that the export memory source as Malloc. + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + armnn::NetworkId networkId = 1; + armnn::INetworkPtr testNetwork(armnn::INetwork::Create()); + + auto inputLayer1 = testNetwork->AddInputLayer(0, "input 1 layer"); + auto inputLayer2 = testNetwork->AddInputLayer(1, "input 2 layer"); + auto addLayer = testNetwork->AddAdditionLayer("add layer"); + auto outputLayer = testNetwork->AddOutputLayer(2, "output layer"); + + TensorInfo tensorInfo{{4}, armnn::DataType::Signed32}; + + inputLayer1->GetOutputSlot(0).Connect(addLayer->GetInputSlot(0)); + inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + inputLayer2->GetOutputSlot(0).Connect(addLayer->GetInputSlot(1)); + inputLayer2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + addLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + addLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + std::vector backends = {armnn::Compute::CpuRef}; + + OptimizerOptions optimizedOptions; + // Hard set import and export to off. + optimizedOptions.m_ImportEnabled = false; + optimizedOptions.m_ExportEnabled = false; + IOptimizedNetworkPtr optNet = Optimize(*testNetwork, backends, runtime->GetDeviceSpec(), optimizedOptions); + CHECK(optNet); + + std::string er; + // Load the network passing an import memory source. + armnn::INetworkProperties networkProperties1(true, MemorySource::Undefined, MemorySource::Malloc); + // There should be an InvalidArgumentException. + runtime->LoadNetwork(networkId, std::move(optNet), er, networkProperties1); + CHECK(er.find("However, it was disabled when this network was optimized") != -1); +} + +TEST_CASE("RuntimeOptimizeImportOn_LoadNetworkImportOff") +{ + // In this test case we'll optimize a network with import enabled. Then we'll attempt to load + // that network but specify that the import memory source is Undefined. + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + armnn::NetworkId networkId = 1; + armnn::INetworkPtr testNetwork(armnn::INetwork::Create()); + + auto inputLayer1 = testNetwork->AddInputLayer(0, "input 1 layer"); + auto inputLayer2 = testNetwork->AddInputLayer(1, "input 2 layer"); + auto addLayer = testNetwork->AddAdditionLayer("add layer"); + auto outputLayer = testNetwork->AddOutputLayer(2, "output layer"); + + TensorInfo tensorInfo{{4}, armnn::DataType::Signed32}; + + inputLayer1->GetOutputSlot(0).Connect(addLayer->GetInputSlot(0)); + inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + inputLayer2->GetOutputSlot(0).Connect(addLayer->GetInputSlot(1)); + inputLayer2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + addLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + addLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + std::vector backends = {armnn::Compute::CpuRef}; + + OptimizerOptions optimizedOptions; + // Hard set import and export to off. + optimizedOptions.m_ImportEnabled = true; + optimizedOptions.m_ExportEnabled = false; + IOptimizedNetworkPtr optNet = Optimize(*testNetwork, backends, runtime->GetDeviceSpec(), optimizedOptions); + CHECK(optNet); + + std::string er; + // Load the network passing an import memory source. + armnn::INetworkProperties networkProperties1(true, MemorySource::Undefined, MemorySource::Undefined); + // There should be an InvalidArgumentException. + runtime->LoadNetwork(networkId, std::move(optNet), er, networkProperties1); + CHECK(er.find("However, it was enabled when this network was optimized") != -1); +} + +TEST_CASE("RuntimeOptimizeExportOn_LoadNetworkExportOff") +{ + // In this test case we'll optimize a network with export enabled. Then we'll attempt to load + // that network but specify that the export memory source is Undefined. + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + armnn::NetworkId networkId = 1; + armnn::INetworkPtr testNetwork(armnn::INetwork::Create()); + + auto inputLayer1 = testNetwork->AddInputLayer(0, "input 1 layer"); + auto inputLayer2 = testNetwork->AddInputLayer(1, "input 2 layer"); + auto addLayer = testNetwork->AddAdditionLayer("add layer"); + auto outputLayer = testNetwork->AddOutputLayer(2, "output layer"); + + TensorInfo tensorInfo{{4}, armnn::DataType::Signed32}; + + inputLayer1->GetOutputSlot(0).Connect(addLayer->GetInputSlot(0)); + inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + inputLayer2->GetOutputSlot(0).Connect(addLayer->GetInputSlot(1)); + inputLayer2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + addLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + addLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + std::vector backends = {armnn::Compute::CpuRef}; + + OptimizerOptions optimizedOptions; + // Hard set import and export to off. + optimizedOptions.m_ImportEnabled = false; + optimizedOptions.m_ExportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*testNetwork, backends, runtime->GetDeviceSpec(), optimizedOptions); + CHECK(optNet); + + std::string er; + // Load the network passing an import memory source. + armnn::INetworkProperties networkProperties1(true, MemorySource::Undefined, MemorySource::Undefined); + // There should be an InvalidArgumentException. + runtime->LoadNetwork(networkId, std::move(optNet), er, networkProperties1); + CHECK(er.find("However, it was enabled when this network was optimized") != -1); +} + } diff --git a/src/armnn/test/TensorHandleStrategyTest.cpp b/src/armnn/test/TensorHandleStrategyTest.cpp index c591fffa43..2ea3c2abf1 100644 --- a/src/armnn/test/TensorHandleStrategyTest.cpp +++ b/src/armnn/test/TensorHandleStrategyTest.cpp @@ -342,7 +342,7 @@ TEST_CASE("TensorHandleSelectionStrategy") graph.TopologicalSort(); std::vector errors; - auto result = SelectTensorHandleStrategy(graph, backends, registry, true, errors); + auto result = SelectTensorHandleStrategy(graph, backends, registry, true, true, errors); CHECK(result.m_Error == false); CHECK(result.m_Warning == false); diff --git a/src/backends/backendsCommon/test/CompatibilityTests.cpp b/src/backends/backendsCommon/test/CompatibilityTests.cpp index c69a4b5f91..9c85ffcfc3 100644 --- a/src/backends/backendsCommon/test/CompatibilityTests.cpp +++ b/src/backends/backendsCommon/test/CompatibilityTests.cpp @@ -73,7 +73,7 @@ TEST_CASE("Neon_Cl_DirectCompatibility_Test") graph.TopologicalSort(); std::vector errors; - auto result = SelectTensorHandleStrategy(graph, backends, registry, true, errors); + auto result = SelectTensorHandleStrategy(graph, backends, registry, true, true, errors); CHECK(result.m_Error == false); CHECK(result.m_Warning == false); diff --git a/src/backends/backendsCommon/test/EndToEndTestImpl.hpp b/src/backends/backendsCommon/test/EndToEndTestImpl.hpp index 77901df444..cc5aa23ca3 100644 --- a/src/backends/backendsCommon/test/EndToEndTestImpl.hpp +++ b/src/backends/backendsCommon/test/EndToEndTestImpl.hpp @@ -204,7 +204,9 @@ inline void ImportNonAlignedInputPointerTest(std::vector backends) pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32)); // Optimize the network - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ImportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); CHECK(optNet); // Loads it into the runtime. @@ -269,7 +271,10 @@ inline void ExportNonAlignedOutputPointerTest(std::vector backends) pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32)); // Optimize the network - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ImportEnabled = true; + optimizedOptions.m_ExportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); CHECK(optNet); // Loads it into the runtime. @@ -340,7 +345,10 @@ inline void ImportAlignedPointerTest(std::vector backends) pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32)); // Optimize the network - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ImportEnabled = true; + optimizedOptions.m_ExportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); CHECK(optNet); // Loads it into the runtime. @@ -424,7 +432,9 @@ inline void ImportOnlyWorkload(std::vector backends) pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32)); // optimize the network - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ImportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); INFO("Load Network"); // Load it into the runtime. It should pass. @@ -514,7 +524,9 @@ inline void ExportOnlyWorkload(std::vector backends) pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32)); // optimize the network - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ExportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); INFO("Load Network"); // Load it into the runtime. It should pass. @@ -601,7 +613,10 @@ inline void ImportAndExportWorkload(std::vector backends) input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32, 0.0f, 0, true)); pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 4 }, DataType::Float32)); - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ImportEnabled = true; + optimizedOptions.m_ExportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); INFO("Load Network"); // Load it into the runtime. It should pass. @@ -694,7 +709,10 @@ inline void ExportOutputWithSeveralOutputSlotConnectionsTest(std::vectorGetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32)); // Optimize the network - IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + OptimizerOptions optimizedOptions; + optimizedOptions.m_ImportEnabled = true; + optimizedOptions.m_ExportEnabled = true; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizedOptions); // Loads it into the runtime. NetworkId netId; diff --git a/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp b/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp index bcea0610db..cd865def71 100644 --- a/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp +++ b/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp @@ -421,7 +421,7 @@ TEST_CASE("OptimizeNetworkCopy") std::vector preferredBackends { "CpuRef" }; armnn::ModelOptions modelOptions; - armnn::OptimizerOptions optimizerOptions(false, false, false, false, modelOptions); + armnn::OptimizerOptions optimizerOptions(false, false, false, false, modelOptions, false); std::vector errorMessages; // optimize the network. diff --git a/src/backends/cl/test/ClCustomAllocatorTests.cpp b/src/backends/cl/test/ClCustomAllocatorTests.cpp index 139e688dc2..251c98fcad 100644 --- a/src/backends/cl/test/ClCustomAllocatorTests.cpp +++ b/src/backends/cl/test/ClCustomAllocatorTests.cpp @@ -120,6 +120,7 @@ TEST_CASE("ClCustomAllocatorTest") // Optimise ArmNN network OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; armnn::IOptimizedNetworkPtr optNet = Optimize(*myNetwork, {"GpuAcc"}, run->GetDeviceSpec(), optOptions); CHECK(optNet); diff --git a/src/backends/cl/test/ClFallbackTests.cpp b/src/backends/cl/test/ClFallbackTests.cpp index 6ac94337ba..51a983a681 100644 --- a/src/backends/cl/test/ClFallbackTests.cpp +++ b/src/backends/cl/test/ClFallbackTests.cpp @@ -50,6 +50,7 @@ TEST_CASE("ClImportEnabledFallbackToNeon") // optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); @@ -330,6 +331,7 @@ TEST_CASE("ClImportEnabledFallbackSubgraphToNeon") // optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); diff --git a/src/backends/cl/test/ClImportTensorHandleTests.cpp b/src/backends/cl/test/ClImportTensorHandleTests.cpp index 20537b3c81..9a075d2b7d 100644 --- a/src/backends/cl/test/ClImportTensorHandleTests.cpp +++ b/src/backends/cl/test/ClImportTensorHandleTests.cpp @@ -142,6 +142,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClImportEndToEnd") // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); @@ -338,6 +339,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConv2dEndToEnd") // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = false; + optOptions.m_ExportEnabled = false; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); @@ -470,6 +472,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConvertFp16toFp32EndToE // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = false; + optOptions.m_ExportEnabled = false; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(network.GetGraph(), backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); @@ -613,6 +616,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConvertFp32toFp16EndToE // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = false; + optOptions.m_ExportEnabled = false; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(network.GetGraph(), backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); @@ -747,6 +751,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportSimpleConvertFp32toFp16 // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = false; + optOptions.m_ExportEnabled = false; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(network.GetGraph(), backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); @@ -896,6 +901,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportRepeatedInferencesEndTo // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = false; + optOptions.m_ExportEnabled = false; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); @@ -1117,6 +1123,7 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportRepeatedInferencesInver // Optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = false; + optOptions.m_ExportEnabled = false; std::vector backends = {armnn::Compute::GpuAcc}; IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optOptions); CHECK(optNet); diff --git a/src/backends/cl/test/ClOptimizedNetworkTests.cpp b/src/backends/cl/test/ClOptimizedNetworkTests.cpp index cf17eae208..6648759a9a 100644 --- a/src/backends/cl/test/ClOptimizedNetworkTests.cpp +++ b/src/backends/cl/test/ClOptimizedNetworkTests.cpp @@ -130,7 +130,7 @@ TEST_CASE("FastMathEnabledTestOnGpuAcc") auto modelOptionsOut = GetModelOptionsForTesting(optimizedNet.get()); - CHECK(modelOptionsOut.size() == 1); + CHECK(modelOptionsOut.size() == 2); // FastMathEnabled and the Global to hold the import export values. CHECK(modelOptionsOut[0].GetOption(0).GetName() == "FastMathEnabled"); CHECK(modelOptionsOut[0].GetOption(0).GetValue().AsBool() == true); } diff --git a/src/backends/neon/test/NeonFallbackTests.cpp b/src/backends/neon/test/NeonFallbackTests.cpp index d2de843fd9..8e0e0ab99b 100644 --- a/src/backends/neon/test/NeonFallbackTests.cpp +++ b/src/backends/neon/test/NeonFallbackTests.cpp @@ -60,6 +60,7 @@ TEST_CASE("FallbackImportToCpuAcc") std::vector backends = { "MockRef", Compute::CpuAcc }; OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); @@ -203,6 +204,7 @@ TEST_CASE("FallbackPaddingCopyToCpuAcc") std::vector backends = { "MockRef", Compute::CpuAcc }; OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); @@ -338,6 +340,7 @@ TEST_CASE("FallbackImportFromCpuAcc") std::vector backends = { "MockRef", Compute::CpuAcc }; OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); @@ -482,6 +485,7 @@ TEST_CASE("FallbackPaddingCopyFromCpuAcc") std::vector backends = { "MockRef", Compute::CpuAcc }; OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); @@ -746,6 +750,7 @@ TEST_CASE("NeonImportEnabledFallbackToCl") // optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); @@ -1037,6 +1042,7 @@ TEST_CASE("NeonImportEnabledFallbackSubgraphToCl") // optimize the network OptimizerOptions optOptions; optOptions.m_ImportEnabled = true; + optOptions.m_ExportEnabled = true; IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); Graph& graph = GetGraphForTesting(optNet.get()); diff --git a/src/backends/neon/test/NeonOptimizedNetworkTests.cpp b/src/backends/neon/test/NeonOptimizedNetworkTests.cpp index 9b448b270d..dcda9bfd07 100644 --- a/src/backends/neon/test/NeonOptimizedNetworkTests.cpp +++ b/src/backends/neon/test/NeonOptimizedNetworkTests.cpp @@ -106,7 +106,7 @@ TEST_CASE("FastMathEnabledTestOnCpuAcc") auto modelOptionsOut = GetModelOptionsForTesting(optimizedNet.get()); - CHECK(modelOptionsOut.size() == 1); + CHECK(modelOptionsOut.size() == 2); // FastMathEnabled and the Global to hold the import export values. CHECK(modelOptionsOut[0].GetOption(0).GetName() == "FastMathEnabled"); CHECK(modelOptionsOut[0].GetOption(0).GetValue().AsBool() == true); } -- cgit v1.2.1