• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "NeonFullyConnectedWorkload.hpp"
7 
8 #include "NeonWorkloadUtils.hpp"
9 
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 #include <aclCommon/ArmComputeUtils.hpp>
12 
13 #include <armnn/utility/PolymorphicDowncast.hpp>
14 
15 #include <backendsCommon/CpuTensorHandle.hpp>
16 
17 #include <arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h>
18 
19 namespace armnn
20 {
21 using namespace armcomputetensorutils;
22 
NeonFullyConnectedWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const TensorInfo & weights,const TensorInfo & biases,const FullyConnectedDescriptor & descriptor,const ActivationDescriptor * activationDescriptor)23 arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
24                                                        const TensorInfo& output,
25                                                        const TensorInfo& weights,
26                                                        const TensorInfo& biases,
27                                                        const FullyConnectedDescriptor& descriptor,
28                                                        const ActivationDescriptor* activationDescriptor)
29 {
30     if (activationDescriptor)
31     {
32         std::vector<ActivationFunction> activations = {ActivationFunction::ReLu, ActivationFunction::BoundedReLu};
33         if (std::find(activations.begin(), activations.end(), activationDescriptor->m_Function) == activations.end())
34         {
35             return arm_compute::Status{
36                 arm_compute::ErrorCode::RUNTIME_ERROR, "NeonFullyConnectedWorkload :Unsupported Activation Function"};
37         }
38     }
39 
40     const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
41     const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
42     const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
43 
44     arm_compute::TensorInfo aclBiases;
45     arm_compute::TensorInfo *optionalAclBiases = nullptr;
46     if (descriptor.m_BiasEnabled)
47     {
48         aclBiases  = BuildArmComputeTensorInfo(biases);
49         optionalAclBiases = &aclBiases;
50     }
51 
52     const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
53         ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor, activationDescriptor);
54 
55     return arm_compute::NEFullyConnectedLayer::validate(&aclInput,
56                                                         &aclWeights,
57                                                         optionalAclBiases,
58                                                         &aclOutput,
59                                                         fullyConnectedLayerInfo);
60 }
61 
NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager)62 NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
63     const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
64     : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
65 {
66     m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);
67 
68     arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
69     arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
70 
71     m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
72     BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());
73 
74     if (m_Data.m_Parameters.m_BiasEnabled)
75     {
76         m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
77         BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
78     }
79 
80     const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
81 
82     arm_compute::FullyConnectedLayerInfo fc_info  =
83             ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor.m_Parameters, activationInfo);
84 
85     auto layer = std::make_unique<arm_compute::NEFullyConnectedLayer>(memoryManager);
86     layer->configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
87     m_FullyConnectedLayer.reset(layer.release());
88 
89     // Allocate
90     if (m_Data.m_Weight->GetTensorInfo().GetDataType() == DataType::QAsymmU8)
91     {
92         InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
93     }
94     else
95     {
96         InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
97     }
98 
99     if (m_BiasesTensor)
100     {
101         if (m_Data.m_Bias->GetTensorInfo().GetDataType() == DataType::Signed32)
102         {
103             InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
104         }
105         else
106         {
107             InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
108         }
109     }
110 
111     // Force Compute Library to perform the necessary copying and reshaping, after which
112     // delete all the input tensors that will no longer be needed
113     m_FullyConnectedLayer->prepare();
114     FreeUnusedTensors();
115 }
116 
Execute() const117 void NeonFullyConnectedWorkload::Execute() const
118 {
119     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonFullyConnectedWorkload_Execute");
120     m_FullyConnectedLayer->run();
121 }
122 
FreeUnusedTensors()123 void NeonFullyConnectedWorkload::FreeUnusedTensors()
124 {
125     FreeTensorIfUnused(m_WeightsTensor);
126     FreeTensorIfUnused(m_BiasesTensor);
127 }
128 
129 } //namespace armnn
130