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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
|
//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonFullyConnectedWorkload.hpp"
#include "NeonWorkloadUtils.hpp"
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <armnn/backends/TensorHandle.hpp>
#include <arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h>
namespace armnn
{
using namespace armcomputetensorutils;
using ACLMemManagerOnDemand = std::shared_ptr<arm_compute::MemoryManagerOnDemand>;
arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const TensorInfo& weights,
const Optional<TensorInfo>& biases,
const FullyConnectedDescriptor& descriptor,
const ActivationDescriptor* activationDescriptor)
{
// The NEON implemented workload does support both const and non const
// weights. However, in the case of non const weights we'd have to call
// prepare or configure for each inference which we're not setup to do just yet.
if (!weights.IsConstant())
{
return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
"Arm NN NeonFullyConnectedWorkload does not support non constant weights."};
}
const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
aclWeights.set_are_values_constant(weights.IsConstant());
arm_compute::TensorInfo aclBiases;
arm_compute::TensorInfo* optionalAclBiases = nullptr;
if (descriptor.m_BiasEnabled)
{
ARMNN_ASSERT(biases.has_value());
// Same for bias as weights. We don't currently support non const.
if (!biases.value().IsConstant())
{
return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
"Arm NN NeonFullyConnectedWorkload does not support non constant bias."};
}
aclBiases = BuildArmComputeTensorInfo(biases.value());
aclBiases.set_are_values_constant(biases.value().IsConstant());
optionalAclBiases = &aclBiases;
}
const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor, activationDescriptor);
return arm_compute::NEFullyConnectedLayer::validate(&aclInput,
&aclWeights,
optionalAclBiases,
&aclOutput,
fullyConnectedLayerInfo);
}
NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
const WorkloadInfo& info,
ACLMemManagerOnDemand& memoryManager)
: NeonBaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
{
m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);
arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
// Copy the weights' tensor into arm_compute tensor.
m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());
InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
if (m_Data.m_Parameters.m_BiasEnabled)
{
// Copy the biases tensor into arm_compute tensor.
m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
}
const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
arm_compute::FullyConnectedLayerInfo fc_info =
ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor.m_Parameters, activationInfo);
auto layer = std::make_unique<arm_compute::NEFullyConnectedLayer>(memoryManager);
layer->configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
m_FullyConnectedLayer.reset(layer.release());
// Add details for profiling output
WorkloadInfo detailsInfo;
detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo());
if (descriptor.m_Parameters.m_BiasEnabled)
{
detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo());
}
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonFullyConnectedWorkload_Construct",
descriptor.m_Parameters,
detailsInfo,
this->GetGuid());
// Force Compute Library to perform the necessary copying and reshaping.
m_FullyConnectedLayer->prepare();
FreeTensorIfUnused(m_WeightsTensor);
FreeTensorIfUnused(m_BiasesTensor);
}
void NeonFullyConnectedWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonFullyConnectedWorkload_Execute", this->GetGuid());
m_FullyConnectedLayer->run();
}
} //namespace armnn
|
