1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
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
|
