IVGCVSW-1927 Add Neon 8-bit FullyConnected support
Change-Id: Idf4cc7a9a7d3261b9eceb653b999257506cdae76
diff --git a/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.cpp b/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.cpp
new file mode 100644
index 0000000..8cebb4f
--- /dev/null
+++ b/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.cpp
@@ -0,0 +1,110 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonFullyConnectedWorkload.hpp"
+
+#include <backends/aclCommon/ArmComputeTensorUtils.hpp>
+#include <backends/aclCommon/ArmComputeUtils.hpp>
+#include <backends/CpuTensorHandle.hpp>
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
+ const TensorInfo& output,
+ const TensorInfo& weights,
+ const TensorInfo& biases,
+ const FullyConnectedDescriptor& descriptor)
+{
+ const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
+ const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
+ const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
+
+ arm_compute::TensorInfo aclBiases;
+ arm_compute::TensorInfo *optionalAclBiases = nullptr;
+ if (descriptor.m_BiasEnabled)
+ {
+ aclBiases = BuildArmComputeTensorInfo(biases);
+ optionalAclBiases = &aclBiases;
+ }
+
+ const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
+ ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor);
+
+
+ return arm_compute::NEFullyConnectedLayer::validate(&aclInput,
+ &aclWeights,
+ optionalAclBiases,
+ &aclOutput,
+ fullyConnectedLayerInfo);
+}
+
+NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
+ const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
+ : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
+ , m_FullyConnectedLayer(memoryManager)
+{
+ m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);
+
+ arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+ arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+ m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());
+
+ if (m_Data.m_Parameters.m_BiasEnabled)
+ {
+ m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
+ }
+
+ // Construct
+ arm_compute::FullyConnectedLayerInfo fc_info;
+ fc_info.transpose_weights = m_Data.m_Parameters.m_TransposeWeightMatrix;
+ m_FullyConnectedLayer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
+
+ // Allocate
+ if (m_Data.m_Weight->GetTensorInfo().GetDataType() == DataType::QuantisedAsymm8)
+ {
+ InitialiseArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight->GetConstTensor<uint8_t>());
+ }
+ else
+ {
+ InitializeArmComputeTensorDataForFloatTypes(*m_WeightsTensor, m_Data.m_Weight);
+ }
+
+ if (m_BiasesTensor)
+ {
+ if (m_Data.m_Bias->GetTensorInfo().GetDataType() == DataType::Signed32)
+ {
+ InitialiseArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias->GetConstTensor<int32_t>());
+ }
+ else
+ {
+ InitializeArmComputeTensorDataForFloatTypes(*m_BiasesTensor, m_Data.m_Bias);
+ }
+ }
+
+ // Force Compute Library to perform the necessary copying and reshaping, after which
+ // delete all the input tensors that will no longer be needed
+ m_FullyConnectedLayer.prepare();
+ FreeUnusedTensors();
+}
+
+void NeonFullyConnectedWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonFullyConnectedWorkload_Execute");
+ m_FullyConnectedLayer.run();
+}
+
+void NeonFullyConnectedWorkload::FreeUnusedTensors()
+{
+ FreeTensorIfUnused(m_WeightsTensor);
+ FreeTensorIfUnused(m_BiasesTensor);
+}
+
+} //namespace armnn
+