--- /dev/null
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <armnn/Descriptors.hpp>
+#include <armnn/IRuntime.hpp>
+#include <armnn/INetwork.hpp>
+#include <armnn/Types.hpp>
+#include <Runtime.hpp>
+
+#include <boost/test/unit_test.hpp>
+
+BOOST_AUTO_TEST_SUITE(DebugCallback)
+
+namespace
+{
+
+using namespace armnn;
+
+INetworkPtr CreateSimpleNetwork()
+{
+ INetworkPtr net(INetwork::Create());
+
+ IConnectableLayer* input = net->AddInputLayer(0, "Input");
+
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::ReLu;
+ IConnectableLayer* activationLayer = net->AddActivationLayer(descriptor, "Activation:ReLu");
+
+ IConnectableLayer* output = net->AddOutputLayer(0);
+
+ input->GetOutputSlot(0).Connect(activationLayer->GetInputSlot(0));
+ activationLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+ input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 5 }, DataType::Float32));
+ activationLayer->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 1, 5 }, DataType::Float32));
+
+ return net;
+}
+
+BOOST_AUTO_TEST_CASE(RuntimeRegisterDebugCallback)
+{
+ INetworkPtr net = CreateSimpleNetwork();
+
+ IRuntime::CreationOptions options;
+ IRuntimePtr runtime(IRuntime::Create(options));
+
+ // Optimize the network with debug option
+ OptimizerOptions optimizerOptions(false, true);
+ std::vector<BackendId> backends = { "CpuRef" };
+ IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optimizerOptions);
+
+ NetworkId netId;
+ BOOST_TEST(runtime->LoadNetwork(netId, std::move(optNet)) == Status::Success);
+
+ // Set up callback function
+ int callCount = 0;
+ std::vector<TensorShape> tensorShapes;
+ std::vector<unsigned int> slotIndexes;
+ auto mockCallback = [&](LayerGuid guid, unsigned int slotIndex, ITensorHandle* tensor)
+ {
+ slotIndexes.push_back(slotIndex);
+ tensorShapes.push_back(tensor->GetShape());
+ callCount++;
+ };
+
+ runtime->RegisterDebugCallback(netId, mockCallback);
+
+ std::vector<float> inputData({-2, -1, 0, 1, 2});
+ std::vector<float> outputData(5);
+
+ InputTensors inputTensors
+ {
+ {0, ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())}
+ };
+ OutputTensors outputTensors
+ {
+ {0, Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
+ };
+
+ runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
+
+ // Check that the callback was called twice
+ BOOST_TEST(callCount == 2);
+
+ // Check that tensor handles passed to callback have correct shapes
+ const std::vector<TensorShape> expectedShapes({TensorShape({1, 1, 1, 5}), TensorShape({1, 1, 1, 5})});
+ BOOST_TEST(tensorShapes == expectedShapes);
+
+ // Check that slot indexes passed to callback are correct
+ const std::vector<unsigned int> expectedSlotIndexes({0, 0});
+ BOOST_TEST(slotIndexes == expectedSlotIndexes);
+}
+
+} // anonymous namespace
+
+BOOST_AUTO_TEST_SUITE_END()
#include <boost/numeric/conversion/cast.hpp>
-#include <cstring>
#include <algorithm>
#include <iostream>
template <typename T>
void Debug(const TensorInfo& inputInfo,
- const TensorInfo& outputInfo,
const T* inputData,
- T* outputData,
LayerGuid guid,
const std::string& layerName,
unsigned int slotIndex)
}
std::cout << " }" << std::endl;
-
- std::memcpy(outputData, inputData, inputInfo.GetNumElements()*sizeof(T));
}
template void Debug<float>(const TensorInfo& inputInfo,
- const TensorInfo& outputInfo,
const float* inputData,
- float* outputData,
LayerGuid guid,
const std::string& layerName,
unsigned int slotIndex);
template void Debug<uint8_t>(const TensorInfo& inputInfo,
- const TensorInfo& outputInfo,
const uint8_t* inputData,
- uint8_t* outputData,
LayerGuid guid,
const std::string& layerName,
unsigned int slotIndex);
#include <TypeUtils.hpp>
+#include <cstring>
+
namespace armnn
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, GetName() + "_Execute");
const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]);
- const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);
const T* inputData = GetInputTensorData<T>(0, m_Data);
T* outputData = GetOutputTensorData<T>(0, m_Data);
- Debug(inputInfo, outputInfo, inputData, outputData, m_Data.m_Guid, m_Data.m_LayerName, m_Data.m_SlotIndex);
+ if (m_Callback)
+ {
+ m_Callback(m_Data.m_Guid, m_Data.m_SlotIndex, m_Data.m_Inputs[0]);
+ }
+ else
+ {
+ Debug(inputInfo, inputData, m_Data.m_Guid, m_Data.m_LayerName, m_Data.m_SlotIndex);
+ }
+
+ std::memcpy(outputData, inputData, inputInfo.GetNumElements()*sizeof(T));
+
+}
+
+template<armnn::DataType DataType>
+void RefDebugWorkload<DataType>::RegisterDebugCallback(const DebugCallbackFunction& func)
+{
+ m_Callback = func;
}
template class RefDebugWorkload<DataType::Float32>;
class RefDebugWorkload : public TypedWorkload<DebugQueueDescriptor, DataType>
{
public:
+ RefDebugWorkload(const DebugQueueDescriptor& descriptor, const WorkloadInfo& info)
+ : TypedWorkload<DebugQueueDescriptor, DataType>(descriptor, info)
+ , m_Callback(nullptr) {}
+
static const std::string& GetName()
{
static const std::string name = std::string("RefDebug") + GetDataTypeName(DataType) + "Workload";
using TypedWorkload<DebugQueueDescriptor, DataType>::TypedWorkload;
void Execute() const override;
+
+ void RegisterDebugCallback(const DebugCallbackFunction& func) override;
+
+private:
+ DebugCallbackFunction m_Callback;
};
using RefDebugFloat32Workload = RefDebugWorkload<DataType::Float32>;