From a983e4699082a0b1ef685bab7354f2ad9cd37a44 Mon Sep 17 00:00:00 2001 From: Colm Donelan Date: Wed, 20 May 2020 16:12:19 +0100 Subject: Updating Doxygen documentation for 20.05 release. Change-Id: I4d624343ed5fd6ae269c3d53532903084508fd14 Signed-off-by: Colm Donelan --- ...neon_fully_connected_workload_8cpp_source.xhtml | 146 +++++++++++++++++++++ 1 file changed, 146 insertions(+) create mode 100644 20.05/_neon_fully_connected_workload_8cpp_source.xhtml (limited to '20.05/_neon_fully_connected_workload_8cpp_source.xhtml') diff --git a/20.05/_neon_fully_connected_workload_8cpp_source.xhtml b/20.05/_neon_fully_connected_workload_8cpp_source.xhtml new file mode 100644 index 0000000000..aa6923396f --- /dev/null +++ b/20.05/_neon_fully_connected_workload_8cpp_source.xhtml @@ -0,0 +1,146 @@ + + + + + + + + + + + + + +ArmNN: src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp Source File + + + + + + + + + + + + + + + + +
+
+ + + + ArmNN + + + +
+
+  20.05 +
+
+
+ + + + + + + +
+
+ +
+
+
+ +
+ +
+
+ + +
+ +
+ +
+
+
NeonFullyConnectedWorkload.cpp
+
+
+Go to the documentation of this file.
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 #include <aclCommon/ArmComputeTensorUtils.hpp>
10 #include <aclCommon/ArmComputeUtils.hpp>
11 #include <armnn/utility/PolymorphicDowncast.hpp>
12 #include <backendsCommon/CpuTensorHandle.hpp>
13 
14 #include <arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h>
15 
16 namespace armnn
17 {
18 using namespace armcomputetensorutils;
19 
20 arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
21  const TensorInfo& output,
22  const TensorInfo& weights,
23  const TensorInfo& biases,
24  const FullyConnectedDescriptor& descriptor)
25 {
26  const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
27  const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
28  const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
29 
30  arm_compute::TensorInfo aclBiases;
31  arm_compute::TensorInfo *optionalAclBiases = nullptr;
32  if (descriptor.m_BiasEnabled)
33  {
34  aclBiases = BuildArmComputeTensorInfo(biases);
35  optionalAclBiases = &aclBiases;
36  }
37 
38  const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
39  ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor);
40 
41 
42  return arm_compute::NEFullyConnectedLayer::validate(&aclInput,
43  &aclWeights,
44  optionalAclBiases,
45  &aclOutput,
46  fullyConnectedLayerInfo);
47 }
48 
49 NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
50  const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
51  : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
52 {
53  m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);
54 
55  arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
56  arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
57 
58  m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
59  BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());
60 
61  if (m_Data.m_Parameters.m_BiasEnabled)
62  {
63  m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
64  BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
65  }
66 
67  // Construct
68  arm_compute::FullyConnectedLayerInfo fc_info;
69  fc_info.transpose_weights = m_Data.m_Parameters.m_TransposeWeightMatrix;
70 
71  auto layer = std::make_unique<arm_compute::NEFullyConnectedLayer>(memoryManager);
72  layer->configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
73  m_FullyConnectedLayer.reset(layer.release());
74 
75  // Allocate
76  if (m_Data.m_Weight->GetTensorInfo().GetDataType() == DataType::QAsymmU8)
77  {
78  InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
79  }
80  else
81  {
82  InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
83  }
84 
85  if (m_BiasesTensor)
86  {
87  if (m_Data.m_Bias->GetTensorInfo().GetDataType() == DataType::Signed32)
88  {
89  InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
90  }
91  else
92  {
93  InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
94  }
95  }
96 
97  // Force Compute Library to perform the necessary copying and reshaping, after which
98  // delete all the input tensors that will no longer be needed
99  m_FullyConnectedLayer->prepare();
100  FreeUnusedTensors();
101 }
102 
103 void NeonFullyConnectedWorkload::Execute() const
104 {
105  ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonFullyConnectedWorkload_Execute");
106  m_FullyConnectedLayer->run();
107 }
108 
109 void NeonFullyConnectedWorkload::FreeUnusedTensors()
110 {
111  FreeTensorIfUnused(m_WeightsTensor);
112  FreeTensorIfUnused(m_BiasesTensor);
113 }
114 
115 } //namespace armnn
ArmComputeUtils.hpp
+
armnn::FullyConnectedQueueDescriptor::m_Weight
const ConstCpuTensorHandle * m_Weight
Definition: WorkloadData.hpp:149
+ +
armnn::FullyConnectedDescriptor::m_TransposeWeightMatrix
bool m_TransposeWeightMatrix
Enable/disable transpose weight matrix.
Definition: Descriptors.hpp:396
+ +
armnn::BaseWorkload< FullyConnectedQueueDescriptor >::m_Data
const FullyConnectedQueueDescriptor m_Data
Definition: Workload.hpp:46
+
ARMNN_SCOPED_PROFILING_EVENT_NEON
#define ARMNN_SCOPED_PROFILING_EVENT_NEON(name)
Definition: NeonWorkloadUtils.hpp:17
+
ArmComputeTensorUtils.hpp
+
armnn::NeonFullyConnectedWorkload::NeonFullyConnectedWorkload
NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor &descriptor, const WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager)
Definition: NeonFullyConnectedWorkload.cpp:49
+
armnn::QueueDescriptor::ValidateInputsOutputs
void ValidateInputsOutputs(const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
Definition: WorkloadData.cpp:460
+
armnn
Copyright (c) 2020 ARM Limited.
Definition: 00_introduction.dox:25
+ + +
PolymorphicDowncast.hpp
+ +
armnn::ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo
arm_compute::FullyConnectedLayerInfo ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(const FullyConnectedDescriptor &fullyConnectedDesc)
Definition: ArmComputeUtils.hpp:133
+
armnn::TensorInfo::GetDataType
DataType GetDataType() const
Definition: Tensor.hpp:95
+ +
armnn::FullyConnectedDescriptor
A FullyConnectedDescriptor for the FullyConnectedLayer.
Definition: Descriptors.hpp:381
+
armnn::FullyConnectedDescriptor::m_BiasEnabled
bool m_BiasEnabled
Enable/disable bias.
Definition: Descriptors.hpp:394
+
armnn::Status
Status
enumeration
Definition: Types.hpp:26
+
armnn::NeonFullyConnectedWorkloadValidate
arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const TensorInfo &weights, const TensorInfo &biases, const FullyConnectedDescriptor &descriptor)
Definition: NeonFullyConnectedWorkload.cpp:20
+
NeonWorkloadUtils.hpp
+
NeonFullyConnectedWorkload.hpp
+
armnn::InitializeArmComputeTensorData
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)
Definition: NeonWorkloadUtils.hpp:35
+
CpuTensorHandle.hpp
+
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:31
+ +
armnn::WorkloadInfo
Contains information about inputs and outputs to a layer.
Definition: WorkloadInfo.hpp:16
+
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:30
+
armnn::FullyConnectedQueueDescriptor::m_Bias
const ConstCpuTensorHandle * m_Bias
Definition: WorkloadData.hpp:150
+ +
armnn::ConstCpuTensorHandle::GetTensorInfo
const TensorInfo & GetTensorInfo() const
Definition: CpuTensorHandle.hpp:37
+
+
+ + + + -- cgit v1.2.1