IVGCVSW-6896 Fix pre-import when using sync execute.

* Refactor backend capability checks in LoadedNetwork.
* ImportInputs should check the number of tensors does not exceed the
  number of inputs.
* In EnqueueWorkload the check for for the count of input tensors
  was ignoring pre-imported inputs.
* Added checks to verify ImportInputs/ImportOutputs worked as expected
  in EndToEndTestImpl.
* Improve documentation on ImportInputs/ImportOutputs in IRuntime.hpp.
* Disabled import tests in CL and Neon EndToEndTests that cannot work.

Signed-off-by: Colm Donelan <colm.donelan@arm.com>
Change-Id: Iae4b2644a1c9f01ee72bce1afb211661cc9ae2e3

2 files changed, 87 insertions, 27 deletions

diff --git a/src/armnn/LoadedNetwork.cpp b/src/armnn/LoadedNetwork.cpp
index 5dd7b6cd2a..d55b99e10f 100644
--- a/src/armnn/LoadedNetwork.cpp
+++ b/src/armnn/LoadedNetwork.cpp
@@ -252,27 +252,23 @@ LoadedNetwork::LoadedNetwork(std::unique_ptr<IOptimizedNetwork> net,
 
             IBackendInternal* backend = it.first->second.get();
 
-            if (networkProperties.m_AsyncEnabled &&
-                !HasCapability(BackendOptions::BackendOption{"AsyncExecution", true}, backend->GetCapabilities()))
+            // If we're doing async execution verify that the backend supports it and ExternallyManagedMemory.
+            if (networkProperties.m_AsyncEnabled)
             {
-                std::string er = backend->GetId();
-                er += " does not support AsyncExecution";
-                throw BackendCapabilityException(er);
-            }
-
-            if (networkProperties.m_AsyncEnabled &&
-                !HasCapability(BackendOptions::BackendOption{"ExternallyManagedMemory", true},
+                if (!HasCapability(BackendOptions::BackendOption{"AsyncExecution", true}, backend->GetCapabilities()))
+                {
+                    std::string er = backend->GetId();
+                    er += " does not support AsyncExecution";
+                    throw BackendCapabilityException(er);
+                }
+                if (!HasCapability(BackendOptions::BackendOption{"ExternallyManagedMemory", true},
                 backend->GetCapabilities()))
-            {
-                std::string er = backend->GetId();
-                er += " does not support ExternallyManagedMemory\n";
-                er += "AsyncEnabled networks require all backends to support ExternallyManagedMemory";
-                throw BackendCapabilityException(er);
-            }
-
-            if (HasCapability(BackendOptions::BackendOption{"ExternallyManagedMemory", true},backend->GetCapabilities())
-                && (m_NetworkProperties.m_ExternalMemoryManagementEnabled ||  m_NetworkProperties.m_AsyncEnabled))
-            {
+                {
+                    std::string er = backend->GetId();
+                    er += " does not support ExternallyManagedMemory\n";
+                    er += "AsyncEnabled networks require all backends to support ExternallyManagedMemory";
+                    throw BackendCapabilityException(er);
+                }
                 m_SupportsExternallyManagedMemory[backend->GetId()] = true;
                 useExternalMemoryManager = true;
             }
@@ -864,7 +860,9 @@ Status LoadedNetwork::EnqueueWorkload(const InputTensors& inputTensors,
     // Data that must be kept alive for the entire execution of the workload.
     WorkloadData workloadData(inputTensors, outputTensors);
 
-    if (graph.GetNumInputs() != inputTensors.size())
+    // Input tensors can be provided as parameters or pre imported. Either way the number of
+    // tensors should match the number of inputs.
+    if (graph.GetNumInputs() != (inputTensors.size() + preImportedInputIds.size()))
     {
         throw InvalidArgumentException("Number of inputs provided does not match network.");
     }
@@ -875,11 +873,6 @@ Status LoadedNetwork::EnqueueWorkload(const InputTensors& inputTensors,
         m_InputQueue.clear();
         m_InputQueue.reserve(graph.GetNumInputs());
 
-        if (preImportedInputIds.size() > graph.GetNumInputs())
-        {
-            throw InvalidArgumentException("Invalid number of preImportedInputIds");
-        }
-
         unsigned int inputIndex = 0;
         unsigned int importedInputIdIndex = 0;
         std::sort(preImportedInputIds.begin(), preImportedInputIds.end());
@@ -1437,9 +1430,10 @@ std::vector<ImportedInputId> LoadedNetwork::ImportInputs(const InputTensors& inp
         {
             throw MemoryImportException("ImportInputs: Memory Import failed, NetworkProperties.m_ImportEnabled");
         }
-        if (inputTensors.size() != m_OptimizedNetwork->pOptimizedNetworkImpl->GetGraph().GetNumInputs())
+        // The number of pre imported tensors should not exceed the number of inputs.
+        if (inputTensors.size() > m_OptimizedNetwork->pOptimizedNetworkImpl->GetGraph().GetNumInputs())
         {
-            throw MemoryImportException("ImportInputs: Force Import failed, incorrect number of tensors");
+            throw MemoryImportException("ImportInputs: The number of tensors provided exceeds the number of inputs.");
         }
 
         std::vector<ImportedInputId> importedInputs;
diff --git a/src/armnn/test/RuntimeTests.cpp b/src/armnn/test/RuntimeTests.cpp
index 59f65541b8..e0d3a222fe 100644
--- a/src/armnn/test/RuntimeTests.cpp
+++ b/src/armnn/test/RuntimeTests.cpp
@@ -1466,4 +1466,70 @@ TEST_CASE("RuntimeOptimizeExportOn_LoadNetworkExportOff")
     CHECK(er.find("However, it was enabled when this network was optimized") != -1);
 }
 
+TEST_CASE("SyncExecutePreImportInputsHappyPath")
+{
+    // In this test case we'll mix "Pre Import" and pass by reference tensors as input.
+    //
+    // * Create a small network that takes two inputs.
+    // * Optimize it specifying that the inputs and outputs will not be imported or exported.
+    // * Create some malloc input and output tensors.
+    // * Use ImportInputs to import only one of the two inputs.
+    // * Call EnqueueWorkload passing one input tensor and one reference to a pre-imported tensor.
+
+    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<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+
+    std::string er;
+    armnn::INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined);
+    runtime->LoadNetwork(networkId, Optimize(*testNetwork, backends, runtime->GetDeviceSpec()), er, networkProperties);
+
+    std::vector<int> inputData1(4, 10);
+    std::vector<int> inputData2(4, 20);
+    std::vector<int> output(4);
+
+    ConstTensor inputTensor1({ { 4 }, armnn::DataType::Signed32, 0.0f, 0, true }, inputData1.data());
+    ConstTensor inputTensor2({ { 4 }, armnn::DataType::Signed32, 0.0f, 0, true }, inputData2.data());
+    Tensor outputTensor({ { 4 }, armnn::DataType::Signed32 }, output.data());
+
+    // An extra check here: the number of inputs provided to ImportInputs should not exceed the number of inputs
+    // to the network.
+    CHECK_THROWS_AS(runtime->ImportInputs(networkId, { { 0, inputTensor1 }, { 0, inputTensor1 }, { 0, inputTensor1 } },
+                                          MemorySource::Malloc),
+                    armnn::MemoryImportException);
+
+    // Pre Import one of the two input tensors.
+    std::vector<ImportedOutputId> importedInputVec =
+        runtime->ImportInputs(networkId, { { 0, inputTensor1 } }, MemorySource::Malloc);
+    CHECK(importedInputVec.size() == 1);
+    CHECK(importedInputVec[0] == 0);
+
+    // We've pre-imported tensor 1 and we'll pass tensor 2 by reference.
+    InputTensors inputTensors{ { 1, inputTensor2 } };
+    OutputTensors outputTensors{ { 2, outputTensor } };
+
+    // Do the inference
+    auto ret = runtime->EnqueueWorkload(networkId, inputTensors, outputTensors, importedInputVec,
+                                        std::vector<ImportedOutputId>());
+    REQUIRE(ret == Status::Success);
+}
 }