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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonNormalizationFloatWorkload.hpp"
#include <backends/neon/NeonLayerSupport.hpp>
#include <backends/aclCommon/ArmComputeUtils.hpp>
#include <backends/aclCommon/ArmComputeTensorUtils.hpp>
using namespace armnn::armcomputetensorutils;
namespace armnn
{
namespace
{
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,
std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
: FloatWorkload<NormalizationQueueDescriptor>(descriptor, info)
, m_NormalizationLayer(memoryManager)
{
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 = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(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);
m_NormalizationLayer.configure(&input, &output, normalizationInfo);
}
void NeonNormalizationFloatWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonNormalizationFloatWorkload_Execute");
m_NormalizationLayer.run();
}
} //namespace armnn
|
