17 using namespace armcomputetensorutils;
25 const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
26 const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
27 const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
29 arm_compute::TensorInfo aclBiases;
30 arm_compute::TensorInfo *optionalAclBiases =
nullptr;
33 aclBiases = BuildArmComputeTensorInfo(biases);
34 optionalAclBiases = &aclBiases;
37 const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
40 return arm_compute::CLFullyConnectedLayer::validate(&aclInput,
44 fullyConnectedLayerInfo);
48 const WorkloadInfo&
info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
50 , m_FullyConnectedLayer(memoryManager)
52 m_WeightsTensor = std::make_unique<arm_compute::CLTensor>();
57 m_BiasesTensor = std::make_unique<arm_compute::CLTensor>();
67 arm_compute::FullyConnectedLayerInfo fc_info;
69 m_FullyConnectedLayer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
80 m_FullyConnectedLayer.prepare();
90 void ClFullyConnectedWorkload::FreeUnusedTensors()
92 FreeTensorIfUnused(m_WeightsTensor);
93 FreeTensorIfUnused(m_BiasesTensor);
void Execute() const override
void RunClFunction(arm_compute::IFunction &function, const CheckLocation &location)
const TensorInfo & GetTensorInfo() const
const ConstCpuTensorHandle * m_Weight
void InitializeArmComputeClTensorData(arm_compute::CLTensor &clTensor, const ConstCpuTensorHandle *handle)
arm_compute::FullyConnectedLayerInfo ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(const FullyConnectedDescriptor &fullyConnectedDesc)
LayerDescriptor m_Parameters
const ConstCpuTensorHandle * m_Bias
bool m_BiasEnabled
Enable/disable bias.
arm_compute::Status ClFullyConnectedWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const TensorInfo &weights, const TensorInfo &biases, const FullyConnectedDescriptor &descriptor)
const armnn::FullyConnectedQueueDescriptor m_Data
A FullyConnectedDescriptor for the FullyConnectedLayer.
bool m_TransposeWeightMatrix
Enable/disable transpose weight matrix.
void ValidateInputsOutputs(const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
std::vector< ITensorHandle * > m_Outputs
std::vector< ITensorHandle * > m_Inputs
#define ARMNN_SCOPED_PROFILING_EVENT_CL(name)
ClFullyConnectedWorkload(const armnn::FullyConnectedQueueDescriptor &descriptor, const armnn::WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager)