IVGCVSW-6786 Add import if memory aligned option to ExecuteNetwork

Change-Id: Ib038e7b2616195a64715e3a7126da1368bbca1d3
Signed-off-by: Jim Flynn <jim.flynn@arm.com>

4 files changed, 34 insertions, 6 deletions

diff --git a/tests/ExecuteNetwork/ExecuteNetwork.cpp b/tests/ExecuteNetwork/ExecuteNetwork.cpp
index 1f19584c68..92a2946847 100644
--- a/tests/ExecuteNetwork/ExecuteNetwork.cpp
+++ b/tests/ExecuteNetwork/ExecuteNetwork.cpp
@@ -405,6 +405,7 @@ int MainImpl(const ExecuteNetworkParams& params,
         inferenceModelParams.m_ThreadPoolSize                 = params.m_ThreadPoolSize;
         inferenceModelParams.m_OutputDetailsToStdOut          = params.m_OutputDetailsToStdOut;
         inferenceModelParams.m_OutputDetailsOnlyToStdOut      = params.m_OutputDetailsOnlyToStdOut;
+        inferenceModelParams.m_ImportInputsIfAligned          = params.m_ImportInputsIfAligned;
 
         for(const std::string& inputName: params.m_InputNames)
         {
diff --git a/tests/ExecuteNetwork/ExecuteNetworkParams.hpp b/tests/ExecuteNetwork/ExecuteNetworkParams.hpp
index cb8c2fb386..deaf55f6e5 100644
--- a/tests/ExecuteNetwork/ExecuteNetworkParams.hpp
+++ b/tests/ExecuteNetwork/ExecuteNetworkParams.hpp
@@ -65,6 +65,7 @@ struct ExecuteNetworkParams
     std::string                   m_MLGOTuningFilePath;
     TfLiteExecutor                m_TfLiteExecutor;
     size_t                        m_ThreadPoolSize;
+    bool                          m_ImportInputsIfAligned;
 
     // Ensures that the parameters for ExecuteNetwork fit together
     void ValidateParams();
diff --git a/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp b/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp
index 8ee66cf64b..681dc8a611 100644
--- a/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp
+++ b/tests/ExecuteNetwork/ExecuteNetworkProgramOptions.cpp
@@ -423,8 +423,12 @@ ProgramOptions::ProgramOptions() : m_CxxOptions{"ExecuteNetwork",
                 ("output-network-details-only",
                  "Outputs layer tensor infos and descriptors to std out without profiling events. Defaults to off.",
                  cxxopts::value<bool>(m_ExNetParams.m_OutputDetailsOnlyToStdOut)->default_value("false")
-                                                                                ->implicit_value("true"));
+                                                                                ->implicit_value("true"))
 
+                ("import-inputs-if-aligned",
+                 "In & Out tensors will be imported per inference if the memory alignment allows. Defaults to false.",
+                 cxxopts::value<bool>(m_ExNetParams.m_ImportInputsIfAligned)->default_value("false")
+                         ->implicit_value("true"));
     }
     catch (const std::exception& e)
     {
diff --git a/tests/InferenceModel.hpp b/tests/InferenceModel.hpp
index bddaf557fd..e2a1a97568 100644
--- a/tests/InferenceModel.hpp
+++ b/tests/InferenceModel.hpp
@@ -111,6 +111,7 @@ struct Params
     std::string                     m_MLGOTuningFilePath;
     bool                            m_AsyncEnabled;
     size_t                          m_ThreadPoolSize;
+    bool                            m_ImportInputsIfAligned;
 
 
     Params()
@@ -132,6 +133,7 @@ struct Params
         , m_MLGOTuningFilePath("")
         , m_AsyncEnabled(false)
         , m_ThreadPoolSize(0)
+        , m_ImportInputsIfAligned(false)
     {}
 };
 
@@ -438,8 +440,9 @@ public:
                    const std::string& dynamicBackendsPath,
                    const std::shared_ptr<armnn::IRuntime>& runtime = nullptr)
         : m_EnableProfiling(enableProfiling),
-          m_ProfilingDetailsMethod(armnn::ProfilingDetailsMethod::Undefined)
-        , m_DynamicBackendsPath(dynamicBackendsPath)
+          m_ProfilingDetailsMethod(armnn::ProfilingDetailsMethod::Undefined),
+          m_DynamicBackendsPath(dynamicBackendsPath),
+          m_ImportInputsIfAligned(params.m_ImportInputsIfAligned)
     {
         if (runtime)
         {
@@ -612,9 +615,27 @@ public:
         // Start timer to record inference time in EnqueueWorkload (in milliseconds)
         const auto start_time = armnn::GetTimeNow();
 
-        armnn::Status ret = m_Runtime->EnqueueWorkload(m_NetworkIdentifier,
-                                                       MakeInputTensors(inputContainers),
-                                                       MakeOutputTensors(outputContainers));
+        armnn::Status ret;
+        if (m_ImportInputsIfAligned)
+        {
+            std::vector<armnn::ImportedInputId> importedInputIds = m_Runtime->ImportInputs(
+                m_NetworkIdentifier, MakeInputTensors(inputContainers), armnn::MemorySource::Malloc);
+
+            std::vector<armnn::ImportedOutputId> importedOutputIds = m_Runtime->ImportOutputs(
+                m_NetworkIdentifier, MakeOutputTensors(outputContainers), armnn::MemorySource::Malloc);
+
+            ret = m_Runtime->EnqueueWorkload(m_NetworkIdentifier,
+                                             MakeInputTensors(inputContainers),
+                                             MakeOutputTensors(outputContainers),
+                                             importedInputIds,
+                                             importedOutputIds);
+        }
+        else
+        {
+            ret = m_Runtime->EnqueueWorkload(m_NetworkIdentifier,
+                                             MakeInputTensors(inputContainers),
+                                             MakeOutputTensors(outputContainers));
+        }
         const auto duration = armnn::GetTimeDuration(start_time);
 
         // if profiling is enabled print out the results
@@ -784,6 +805,7 @@ private:
     bool m_EnableProfiling;
     armnn::ProfilingDetailsMethod m_ProfilingDetailsMethod;
     std::string m_DynamicBackendsPath;
+    bool m_ImportInputsIfAligned;
 
     template<typename TContainer>
     armnn::InputTensors MakeInputTensors(const std::vector<TContainer>& inputDataContainers)