aboutsummaryrefslogtreecommitdiff
path: root/src/armnn
diff options
context:
space:
mode:
Diffstat (limited to 'src/armnn')
-rw-r--r--src/armnn/LoadedNetwork.cpp50
-rw-r--r--src/armnn/LoadedNetwork.hpp7
-rw-r--r--src/armnn/Runtime.cpp14
-rw-r--r--src/armnn/Runtime.hpp7
4 files changed, 55 insertions, 23 deletions
diff --git a/src/armnn/LoadedNetwork.cpp b/src/armnn/LoadedNetwork.cpp
index 5b64085869..1000eceda0 100644
--- a/src/armnn/LoadedNetwork.cpp
+++ b/src/armnn/LoadedNetwork.cpp
@@ -41,7 +41,8 @@ std::string ToErrorMessage(const char * prefix, const ExceptionType & error)
} // anonymous
std::unique_ptr<LoadedNetwork> LoadedNetwork::MakeLoadedNetwork(std::unique_ptr<OptimizedNetwork> net,
- std::string & errorMessage)
+ std::string& errorMessage,
+ const INetworkProperties& networkProperties)
{
std::unique_ptr<LoadedNetwork> loadedNetwork;
@@ -55,7 +56,7 @@ std::unique_ptr<LoadedNetwork> LoadedNetwork::MakeLoadedNetwork(std::unique_ptr<
try
{
- loadedNetwork.reset(new LoadedNetwork(std::move(net)));
+ loadedNetwork.reset(new LoadedNetwork(std::move(net), networkProperties));
}
catch (const armnn::RuntimeException& error)
{
@@ -73,8 +74,11 @@ std::unique_ptr<LoadedNetwork> LoadedNetwork::MakeLoadedNetwork(std::unique_ptr<
return loadedNetwork;
}
-LoadedNetwork::LoadedNetwork(std::unique_ptr<OptimizedNetwork> net)
- : m_OptimizedNetwork(std::move(net))
+LoadedNetwork::LoadedNetwork(std::unique_ptr<OptimizedNetwork> net,
+ const INetworkProperties& networkProperties) :
+ m_OptimizedNetwork(std::move(net)),
+ m_IsImportEnabled(networkProperties.m_ImportEnabled),
+ m_IsExportEnabled(networkProperties.m_ExportEnabled)
{
// Create a profiler and register it for the current thread.
m_Profiler = std::make_shared<Profiler>();
@@ -392,7 +396,7 @@ void LoadedNetwork::EnqueueInput(const BindableLayer& layer, ITensorHandle* tens
info.m_OutputTensorInfos.push_back(outputTensorInfo);
MemorySourceFlags importFlags = outputTensorHandle->GetImportFlags();
- if (CheckFlag(importFlags, MemorySource::Malloc)) // Try import the input tensor
+ if (CheckFlag(importFlags, MemorySource::Malloc) && m_IsImportEnabled) // Try import the input tensor
{
// This assumes a CPU Tensor handle
void* mem = tensorHandle->Map(false);
@@ -402,13 +406,16 @@ void LoadedNetwork::EnqueueInput(const BindableLayer& layer, ITensorHandle* tens
return; // No need for a workload since the import has been done.
}
tensorHandle->Unmap();
+ throw MemoryImportException("EnqueueInput: Memory Import failed");
}
+ else
+ {
+ // Create a mem copy workload for input since we did not import
+ auto inputWorkload = std::make_unique<CopyMemGenericWorkload>(inputQueueDescriptor, info);
- // Create a mem copy workload for input since we could not import
- auto inputWorkload = std::make_unique<CopyMemGenericWorkload>(inputQueueDescriptor, info);
-
- BOOST_ASSERT_MSG(inputWorkload, "No input workload created");
- m_InputQueue.push_back(move(inputWorkload));
+ BOOST_ASSERT_MSG(inputWorkload, "No input workload created");
+ m_InputQueue.push_back(move(inputWorkload));
+ }
}
void LoadedNetwork::EnqueueOutput(const BindableLayer& layer, ITensorHandle* tensorHandle, const TensorInfo& tensorInfo)
@@ -444,7 +451,8 @@ void LoadedNetwork::EnqueueOutput(const BindableLayer& layer, ITensorHandle* ten
// b) The tensor has zero padding
// c) There is only one connection to the OutputSlot and it is to an OutputLayer.
// d) The output pointer is allocated via malloc. (Other types will be supported in a later release)
- if (layer.GetInputSlots()[0].GetConnectedOutputSlot()->GetOwningLayer().GetType() != LayerType::Input)
+ if (layer.GetInputSlots()[0].GetConnectedOutputSlot()->GetOwningLayer().GetType() != LayerType::Input
+ && m_IsExportEnabled)
{
if (layer.GetInputSlots()[0].GetConnectedOutputSlot()->GetNumConnections() == 1)
{
@@ -467,17 +475,23 @@ void LoadedNetwork::EnqueueOutput(const BindableLayer& layer, ITensorHandle* ten
return; //No need to add the output workload below
}
+ else
+ {
+ throw MemoryExportException("EnqueueOutput: Memory Export failed");
+ }
}
}
}
+ else
+ {
+ // If we got here then we couldn't import the memory, so add an output workload which performs a memcopy.
+ outputQueueDescriptor.m_Inputs.push_back(inputTensorHandle);
+ info.m_InputTensorInfos.push_back(inputTensorInfo);
- // If we got here then we couldn't import the memory, so add an output workload which performs a memcopy.
- outputQueueDescriptor.m_Inputs.push_back(inputTensorHandle);
- info.m_InputTensorInfos.push_back(inputTensorInfo);
-
- auto outputWorkload = std::make_unique<CopyMemGenericWorkload>(outputQueueDescriptor, info);
- BOOST_ASSERT_MSG(outputWorkload, "No output workload created");
- m_OutputQueue.push_back(move(outputWorkload));
+ auto outputWorkload = std::make_unique<CopyMemGenericWorkload>(outputQueueDescriptor, info);
+ BOOST_ASSERT_MSG(outputWorkload, "No output workload created");
+ m_OutputQueue.push_back(move(outputWorkload));
+ }
}
void LoadedNetwork::AllocateWorkingMemory()
diff --git a/src/armnn/LoadedNetwork.hpp b/src/armnn/LoadedNetwork.hpp
index 808a93222a..08c09b8801 100644
--- a/src/armnn/LoadedNetwork.hpp
+++ b/src/armnn/LoadedNetwork.hpp
@@ -41,7 +41,8 @@ public:
Status EnqueueWorkload(const InputTensors& inputTensors, const OutputTensors& outputTensors);
static std::unique_ptr<LoadedNetwork> MakeLoadedNetwork(std::unique_ptr<OptimizedNetwork> net,
- std::string & errorMessage);
+ std::string & errorMessage,
+ const INetworkProperties& networkProperties);
// NOTE we return by reference as the purpose of this method is only to provide
// access to the private m_Profiler and in theory we should not need to increment
@@ -55,7 +56,7 @@ public:
private:
void AllocateWorkingMemory();
- LoadedNetwork(std::unique_ptr<OptimizedNetwork> net);
+ LoadedNetwork(std::unique_ptr<OptimizedNetwork> net, const INetworkProperties& networkProperties);
void EnqueueInput(const BindableLayer& layer, ITensorHandle* tensorHandle, const TensorInfo& tensorInfo);
@@ -84,6 +85,8 @@ private:
mutable std::mutex m_WorkingMemMutex;
bool m_IsWorkingMemAllocated=false;
+ bool m_IsImportEnabled=false;
+ bool m_IsExportEnabled=false;
TensorHandleFactoryRegistry m_TensorHandleFactoryRegistry;
};
diff --git a/src/armnn/Runtime.cpp b/src/armnn/Runtime.cpp
index 9e874848ec..e47835687d 100644
--- a/src/armnn/Runtime.cpp
+++ b/src/armnn/Runtime.cpp
@@ -49,7 +49,16 @@ Status Runtime::LoadNetwork(NetworkId& networkIdOut, IOptimizedNetworkPtr inNetw
Status Runtime::LoadNetwork(NetworkId& networkIdOut,
IOptimizedNetworkPtr inNetwork,
- std::string & errorMessage)
+ std::string& errorMessage)
+{
+ INetworkProperties networkProperties;
+ return LoadNetwork(networkIdOut, std::move(inNetwork), errorMessage, networkProperties);
+}
+
+Status Runtime::LoadNetwork(NetworkId& networkIdOut,
+ IOptimizedNetworkPtr inNetwork,
+ std::string& errorMessage,
+ const INetworkProperties& networkProperties)
{
IOptimizedNetwork* rawNetwork = inNetwork.release();
@@ -62,7 +71,8 @@ Status Runtime::LoadNetwork(NetworkId& networkIdOut,
unique_ptr<LoadedNetwork> loadedNetwork = LoadedNetwork::MakeLoadedNetwork(
std::unique_ptr<OptimizedNetwork>(boost::polymorphic_downcast<OptimizedNetwork*>(rawNetwork)),
- errorMessage);
+ errorMessage,
+ networkProperties);
if (!loadedNetwork)
{
diff --git a/src/armnn/Runtime.hpp b/src/armnn/Runtime.hpp
index 35684f1f78..a028c878a0 100644
--- a/src/armnn/Runtime.hpp
+++ b/src/armnn/Runtime.hpp
@@ -38,7 +38,12 @@ public:
/// @return armnn::Status
virtual Status LoadNetwork(NetworkId& networkIdOut,
IOptimizedNetworkPtr network,
- std::string & errorMessage) override;
+ std::string& errorMessage) override;
+
+ virtual Status LoadNetwork(NetworkId& networkIdOut,
+ IOptimizedNetworkPtr network,
+ std::string& errorMessage,
+ const INetworkProperties& networkProperties) override;
virtual TensorInfo GetInputTensorInfo(NetworkId networkId, LayerBindingId layerId) const override;
virtual TensorInfo GetOutputTensorInfo(NetworkId networkId, LayerBindingId layerId) const override;
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-18 17:10:42 +0000