1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonBatchNormalizationWorkload.hpp"
7
8 #include "NeonWorkloadUtils.hpp"
9
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 #include <aclCommon/ArmComputeUtils.hpp>
12
13 #include <armnn/utility/PolymorphicDowncast.hpp>
14
15 #include <armnn/backends/TensorHandle.hpp>
16
17 #include <arm_compute/runtime/NEON/functions/NEBatchNormalizationLayer.h>
18
19 namespace armnn
20 {
21 using namespace armcomputetensorutils;
22
23
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)24 arm_compute::Status NeonBatchNormalizationValidate(const TensorInfo& input,
25 const TensorInfo& output,
26 const TensorInfo& mean,
27 const TensorInfo& var,
28 const TensorInfo& beta,
29 const TensorInfo& gamma,
30 const BatchNormalizationDescriptor& descriptor,
31 const ActivationDescriptor* activationDescriptor)
32 {
33 const arm_compute::TensorInfo aclInputInfo =
34 armcomputetensorutils::BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
35 const arm_compute::TensorInfo aclOutputInfo =
36 armcomputetensorutils::BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
37 const arm_compute::TensorInfo aclMeanInfo =
38 armcomputetensorutils::BuildArmComputeTensorInfo(mean, descriptor.m_DataLayout);
39 const arm_compute::TensorInfo aclVarInfo =
40 armcomputetensorutils::BuildArmComputeTensorInfo(var, descriptor.m_DataLayout);
41 const arm_compute::TensorInfo aclBetaInfo =
42 armcomputetensorutils::BuildArmComputeTensorInfo(beta, descriptor.m_DataLayout);
43 const arm_compute::TensorInfo aclGammaInfo =
44 armcomputetensorutils::BuildArmComputeTensorInfo(gamma, descriptor.m_DataLayout);
45
46 const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
47 activationDescriptor);
48
49 return arm_compute::NEBatchNormalizationLayer::validate(&aclInputInfo,
50 &aclOutputInfo,
51 &aclMeanInfo,
52 &aclVarInfo,
53 &aclBetaInfo,
54 &aclGammaInfo,
55 descriptor.m_Eps,
56 activationInfo);
57 }
58
NeonBatchNormalizationWorkload(const BatchNormalizationQueueDescriptor & descriptor,const WorkloadInfo & info)59 NeonBatchNormalizationWorkload::NeonBatchNormalizationWorkload(
60 const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info)
61 : NeonBaseWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
62 {
63 // Report Profiling Details
64 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonBatchNormalizationWorkload_Construct",
65 descriptor.m_Parameters,
66 info,
67 this->GetGuid());
68
69 m_Data.ValidateInputsOutputs("NeonBatchNormalizationWorkload", 1, 1);
70
71 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
72 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
73
74 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
75 input.info()->set_data_layout(aclDataLayout);
76 output.info()->set_data_layout(aclDataLayout);
77
78 m_Mean = std::make_unique<arm_compute::Tensor>();
79 BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo());
80
81 m_Variance = std::make_unique<arm_compute::Tensor>();
82 BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo());
83
84 m_Gamma = std::make_unique<arm_compute::Tensor>();
85 BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
86
87 m_Beta = std::make_unique<arm_compute::Tensor>();
88 BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo());
89
90 const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
91
92 auto layer = std::make_unique<arm_compute::NEBatchNormalizationLayer>();
93 layer->configure(&input,
94 &output,
95 m_Mean.get(),
96 m_Variance.get(),
97 m_Beta.get(),
98 m_Gamma.get(),
99 m_Data.m_Parameters.m_Eps,
100 activationInfo);
101 m_Layer.reset(layer.release());
102
103 InitializeArmComputeTensorData(*m_Mean, m_Data.m_Mean);
104 InitializeArmComputeTensorData(*m_Variance, m_Data.m_Variance);
105 InitializeArmComputeTensorData(*m_Gamma, m_Data.m_Gamma);
106 InitializeArmComputeTensorData(*m_Beta, m_Data.m_Beta);
107
108 // Force Compute Library to perform the necessary copying and reshaping, after which
109 // delete all the input tensors that will no longer be needed
110 m_Layer->prepare();
111 FreeUnusedTensors();
112 }
113
Execute() const114 void NeonBatchNormalizationWorkload::Execute() const
115 {
116 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonBatchNormalizationWorkload_Execute", this->GetGuid());
117 m_Layer->run();
118 }
119
FreeUnusedTensors()120 void NeonBatchNormalizationWorkload::FreeUnusedTensors()
121 {
122 FreeTensorIfUnused(m_Mean);
123 FreeTensorIfUnused(m_Variance);
124 FreeTensorIfUnused(m_Gamma);
125 FreeTensorIfUnused(m_Beta);
126 }
127
128 } //namespace armnn
129