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
|
//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonBatchNormalizationWorkload.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/NEBatchNormalizationLayer.h>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status NeonBatchNormalizationValidate(const TensorInfo& input,
const TensorInfo& output,
const TensorInfo& mean,
const TensorInfo& var,
const TensorInfo& beta,
const TensorInfo& gamma,
const BatchNormalizationDescriptor& descriptor,
const ActivationDescriptor* activationDescriptor)
{
const arm_compute::TensorInfo aclInputInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclMeanInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(mean, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclVarInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(var, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclBetaInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(beta, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclGammaInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(gamma, descriptor.m_DataLayout);
const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
activationDescriptor);
return arm_compute::NEBatchNormalizationLayer::validate(&aclInputInfo,
&aclOutputInfo,
&aclMeanInfo,
&aclVarInfo,
&aclBetaInfo,
&aclGammaInfo,
descriptor.m_Eps,
activationInfo);
}
NeonBatchNormalizationWorkload::NeonBatchNormalizationWorkload(
const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info)
: NeonBaseWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
{
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonBatchNormalizationWorkload_Construct",
descriptor.m_Parameters,
info,
this->GetGuid());
m_Data.ValidateInputsOutputs("NeonBatchNormalizationWorkload", 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();
arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
input.info()->set_data_layout(aclDataLayout);
output.info()->set_data_layout(aclDataLayout);
m_Mean = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo());
m_Variance = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo());
m_Gamma = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
m_Beta = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo());
const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
auto layer = std::make_unique<arm_compute::NEBatchNormalizationLayer>();
layer->configure(&input,
&output,
m_Mean.get(),
m_Variance.get(),
m_Beta.get(),
m_Gamma.get(),
m_Data.m_Parameters.m_Eps,
activationInfo);
m_Layer.reset(layer.release());
InitializeArmComputeTensorData(*m_Mean, m_Data.m_Mean);
InitializeArmComputeTensorData(*m_Variance, m_Data.m_Variance);
InitializeArmComputeTensorData(*m_Gamma, m_Data.m_Gamma);
InitializeArmComputeTensorData(*m_Beta, m_Data.m_Beta);
// Force Compute Library to perform the necessary copying and reshaping, after which
// delete all the input tensors that will no longer be needed
m_Layer->prepare();
FreeUnusedTensors();
}
void NeonBatchNormalizationWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonBatchNormalizationWorkload_Execute", this->GetGuid());
m_Layer->run();
}
void NeonBatchNormalizationWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_Mean);
FreeTensorIfUnused(m_Variance);
FreeTensorIfUnused(m_Gamma);
FreeTensorIfUnused(m_Beta);
}
} //namespace armnn
|
