1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonNormalizationFloatWorkload.hpp"
7
8 #include "NeonWorkloadUtils.hpp"
9 #include <aclCommon/ArmComputeUtils.hpp>
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 #include <armnn/utility/PolymorphicDowncast.hpp>
12
13 #include <arm_compute/runtime/NEON/functions/NENormalizationLayer.h>
14
15 using namespace armnn::armcomputetensorutils;
16
17 namespace armnn
18 {
19
20 namespace
21 {
22
IsNeonNormalizationDescriptorSupported(const NormalizationDescriptor & parameters,Optional<std::string &> reasonIfUnsupported)23 bool IsNeonNormalizationDescriptorSupported(const NormalizationDescriptor& parameters,
24 Optional<std::string&> reasonIfUnsupported)
25 {
26 if (parameters.m_NormMethodType != NormalizationAlgorithmMethod::LocalBrightness)
27 {
28 if (reasonIfUnsupported)
29 {
30 reasonIfUnsupported.value() = "Unsupported normalisation method type, only LocalBrightness is supported";
31 }
32 return false;
33 }
34 if (parameters.m_NormSize % 2 == 0)
35 {
36 if (reasonIfUnsupported)
37 {
38 reasonIfUnsupported.value() = "Normalization size must be an odd number.";
39 }
40 return false;
41 }
42
43 return true;
44 }
45
46 } // anonymous namespace
47
NeonNormalizationWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const NormalizationDescriptor & descriptor)48 arm_compute::Status NeonNormalizationWorkloadValidate(const TensorInfo& input,
49 const TensorInfo& output,
50 const NormalizationDescriptor& descriptor)
51 {
52 const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
53 const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
54
55 arm_compute::NormalizationLayerInfo normalizationInfo = BuildArmComputeNormalizationLayerInfo(descriptor);
56
57 return arm_compute::NENormalizationLayer::validate(&aclInput, &aclOutput, normalizationInfo);
58 }
59
NeonNormalizationFloatWorkload(const NormalizationQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager)60 NeonNormalizationFloatWorkload::NeonNormalizationFloatWorkload(const NormalizationQueueDescriptor& descriptor,
61 const WorkloadInfo& info,
62 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
63 : FloatWorkload<NormalizationQueueDescriptor>(descriptor, info)
64 {
65 m_Data.ValidateInputsOutputs("NeonNormalizationFloatWorkload", 1, 1);
66 std::string reasonIfUnsupported;
67 if (!IsNeonNormalizationDescriptorSupported(m_Data.m_Parameters, Optional<std::string&>(reasonIfUnsupported)))
68 {
69 throw UnimplementedException(reasonIfUnsupported);
70 }
71
72 // Input and output tensors have to have the same dimensionality.
73 if (info.m_InputTensorInfos[0].GetShape()[1] != info.m_OutputTensorInfos[0].GetShape()[1]
74 || info.m_InputTensorInfos[0].GetShape()[0] != info.m_OutputTensorInfos[0].GetShape()[0]
75 || info.m_InputTensorInfos[0].GetShape()[3] != info.m_OutputTensorInfos[0].GetShape()[3]
76 || info.m_InputTensorInfos[0].GetShape()[2] != info.m_OutputTensorInfos[0].GetShape()[2])
77 {
78 throw InvalidArgumentException("Normalization requires input and output tensors to have equal dimensionality.");
79 }
80
81 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
82 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
83 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
84 input.info()->set_data_layout(aclDataLayout);
85 output.info()->set_data_layout(aclDataLayout);
86
87 const arm_compute::NormType normType =
88 ConvertNormalizationAlgorithmChannelToAclNormType(m_Data.m_Parameters.m_NormChannelType);
89 arm_compute::NormalizationLayerInfo normalizationInfo(normType,
90 m_Data.m_Parameters.m_NormSize,
91 m_Data.m_Parameters.m_Alpha,
92 m_Data.m_Parameters.m_Beta,
93 m_Data.m_Parameters.m_K,
94 false);
95 auto layer = std::make_unique<arm_compute::NENormalizationLayer>(memoryManager);
96 layer->configure(&input, &output, normalizationInfo);
97 m_NormalizationLayer.reset(layer.release());
98 }
99
Execute() const100 void NeonNormalizationFloatWorkload::Execute() const
101 {
102 ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonNormalizationFloatWorkload_Execute");
103 m_NormalizationLayer->run();
104 }
105
106 } //namespace armnn
107