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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
|
//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonNormalizationFloatWorkload.hpp"
#include "NeonWorkloadUtils.hpp"
#include <aclCommon/ArmComputeUtils.hpp>
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <arm_compute/runtime/NEON/functions/NENormalizationLayer.h>
using namespace armnn::armcomputetensorutils;
namespace armnn
{
namespace
{
using ACLMemManagerOnDemand = std::shared_ptr<arm_compute::MemoryManagerOnDemand>;
bool IsNeonNormalizationDescriptorSupported(const NormalizationDescriptor& parameters,
Optional<std::string&> reasonIfUnsupported)
{
if (parameters.m_NormMethodType != NormalizationAlgorithmMethod::LocalBrightness)
{
if (reasonIfUnsupported)
{
reasonIfUnsupported.value() = "Unsupported normalisation method type, only LocalBrightness is supported";
}
return false;
}
if (parameters.m_NormSize % 2 == 0)
{
if (reasonIfUnsupported)
{
reasonIfUnsupported.value() = "Normalization size must be an odd number.";
}
return false;
}
return true;
}
} // anonymous namespace
arm_compute::Status NeonNormalizationWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const NormalizationDescriptor& descriptor)
{
const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
arm_compute::NormalizationLayerInfo normalizationInfo = BuildArmComputeNormalizationLayerInfo(descriptor);
return arm_compute::NENormalizationLayer::validate(&aclInput, &aclOutput, normalizationInfo);
}
NeonNormalizationFloatWorkload::NeonNormalizationFloatWorkload(const NormalizationQueueDescriptor& descriptor,
const WorkloadInfo& info,
ACLMemManagerOnDemand& memoryManager)
: FloatWorkload<NormalizationQueueDescriptor>(descriptor, info)
{
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonNormalizationWorkload_Construct",
descriptor.m_Parameters,
info,
this->GetGuid());
m_Data.ValidateInputsOutputs("NeonNormalizationFloatWorkload", 1, 1);
std::string reasonIfUnsupported;
if (!IsNeonNormalizationDescriptorSupported(m_Data.m_Parameters, Optional<std::string&>(reasonIfUnsupported)))
{
throw UnimplementedException(reasonIfUnsupported);
}
// Input and output tensors have to have the same dimensionality.
if (info.m_InputTensorInfos[0].GetShape()[1] != info.m_OutputTensorInfos[0].GetShape()[1]
|| info.m_InputTensorInfos[0].GetShape()[0] != info.m_OutputTensorInfos[0].GetShape()[0]
|| info.m_InputTensorInfos[0].GetShape()[3] != info.m_OutputTensorInfos[0].GetShape()[3]
|| info.m_InputTensorInfos[0].GetShape()[2] != info.m_OutputTensorInfos[0].GetShape()[2])
{
throw InvalidArgumentException("Normalization requires input and output tensors to have equal dimensionality.");
}
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);
const arm_compute::NormType normType =
ConvertNormalizationAlgorithmChannelToAclNormType(m_Data.m_Parameters.m_NormChannelType);
arm_compute::NormalizationLayerInfo normalizationInfo(normType,
m_Data.m_Parameters.m_NormSize,
m_Data.m_Parameters.m_Alpha,
m_Data.m_Parameters.m_Beta,
m_Data.m_Parameters.m_K,
false);
auto layer = std::make_unique<arm_compute::NENormalizationLayer>(memoryManager);
layer->configure(&input, &output, normalizationInfo);
m_NormalizationLayer.reset(layer.release());
}
void NeonNormalizationFloatWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonNormalizationFloatWorkload_Execute", this->GetGuid());
m_NormalizationLayer->run();
}
void NeonNormalizationFloatWorkload::ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
{
ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
this->m_Data.m_Inputs[slot] = tensorHandle;
try
{
Reconfigure();
}
catch(armnn::UnimplementedException& e)
{
// Cannot reconfigure, revert the slot back and throw the exception.
this->m_Data.m_Inputs[slot] = backupHandle;
throw e;
}
}
// Replace output tensor handle with the given TensorHandle
void NeonNormalizationFloatWorkload::ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
{
ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
this->m_Data.m_Inputs[slot] = tensorHandle;
try
{
Reconfigure();
}
catch(armnn::UnimplementedException& e)
{
// Cannot reconfigure, revert the slot back and throw the exception.
this->m_Data.m_Inputs[slot] = backupHandle;
throw e;
}
}
void NeonNormalizationFloatWorkload::Reconfigure()
{
throw armnn::UnimplementedException("Reconfigure not implemented for this workload");
}
} //namespace armnn
|
