1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ClBatchNormalizationFloatWorkload.hpp"
7 #include "ClWorkloadUtils.hpp"
8
9 #include <aclCommon/ArmComputeTensorUtils.hpp>
10 #include <aclCommon/ArmComputeUtils.hpp>
11 #include <armnn/backends/TensorHandle.hpp>
12 #include <cl/ClLayerSupport.hpp>
13 #include <cl/ClTensorHandle.hpp>
14
15 namespace armnn
16 {
17 using namespace armcomputetensorutils;
18
ClBatchNormalizationValidate(const TensorInfo & input,const TensorInfo & output,const TensorInfo & mean,const TensorInfo & var,const TensorInfo & beta,const TensorInfo & gamma,const BatchNormalizationDescriptor & descriptor,const ActivationDescriptor * activationDescriptor)19 arm_compute::Status ClBatchNormalizationValidate(const TensorInfo& input,
20 const TensorInfo& output,
21 const TensorInfo& mean,
22 const TensorInfo& var,
23 const TensorInfo& beta,
24 const TensorInfo& gamma,
25 const BatchNormalizationDescriptor& descriptor,
26 const ActivationDescriptor* activationDescriptor)
27 {
28 const arm_compute::TensorInfo aclInputInfo =
29 armcomputetensorutils::BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
30 const arm_compute::TensorInfo aclOutputInfo =
31 armcomputetensorutils::BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
32 const arm_compute::TensorInfo aclMeanInfo =
33 armcomputetensorutils::BuildArmComputeTensorInfo(mean, descriptor.m_DataLayout);
34 const arm_compute::TensorInfo aclVarInfo =
35 armcomputetensorutils::BuildArmComputeTensorInfo(var, descriptor.m_DataLayout);
36 const arm_compute::TensorInfo aclBetaInfo =
37 armcomputetensorutils::BuildArmComputeTensorInfo(beta, descriptor.m_DataLayout);
38 const arm_compute::TensorInfo aclGammaInfo =
39 armcomputetensorutils::BuildArmComputeTensorInfo(gamma, descriptor.m_DataLayout);
40
41 const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
42 activationDescriptor);
43
44 return arm_compute::CLBatchNormalizationLayer::validate(&aclInputInfo,
45 &aclOutputInfo,
46 &aclMeanInfo,
47 &aclVarInfo,
48 &aclBetaInfo,
49 &aclGammaInfo,
50 descriptor.m_Eps,
51 activationInfo);
52 }
53
ClBatchNormalizationFloatWorkload(const BatchNormalizationQueueDescriptor & descriptor,const WorkloadInfo & info,const arm_compute::CLCompileContext & clCompileContext)54 ClBatchNormalizationFloatWorkload::ClBatchNormalizationFloatWorkload(
55 const BatchNormalizationQueueDescriptor& descriptor,
56 const WorkloadInfo& info,
57 const arm_compute::CLCompileContext& clCompileContext)
58 : FloatWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
59 {
60 // Report Profiling Details
61 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchNormalizationWorkload_Construct",
62 descriptor.m_Parameters,
63 info,
64 this->GetGuid());
65
66 m_Mean = std::make_unique<arm_compute::CLTensor>();
67 BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo());
68
69 m_Variance = std::make_unique<arm_compute::CLTensor>();
70 BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo());
71
72 m_Gamma = std::make_unique<arm_compute::CLTensor>();
73 BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
74
75 m_Beta = std::make_unique<arm_compute::CLTensor>();
76 BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo());
77
78 m_Data.ValidateInputsOutputs("ClBatchNormalizationFloatWorkload", 1, 1);
79
80 arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
81 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
82
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::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
88
89 {
90 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClBatchNormalizationFloatWorkload_configure");
91 m_Layer.configure(clCompileContext,
92 &input,
93 &output,
94 m_Mean.get(),
95 m_Variance.get(),
96 m_Beta.get(),
97 m_Gamma.get(),
98 m_Data.m_Parameters.m_Eps,
99 activationInfo);
100 }
101
102 InitializeArmComputeClTensorData(*m_Mean, m_Data.m_Mean);
103 InitializeArmComputeClTensorData(*m_Variance, m_Data.m_Variance);
104 InitializeArmComputeClTensorData(*m_Beta, m_Data.m_Beta);
105 InitializeArmComputeClTensorData(*m_Gamma, m_Data.m_Gamma);
106
107 // Force Compute Library to perform the necessary copying and reshaping, after which
108 // delete all the input tensors that will no longer be needed
109 m_Layer.prepare();
110 FreeUnusedTensors();
111 }
112
Execute() const113 void ClBatchNormalizationFloatWorkload::Execute() const
114 {
115 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClBatchNormalizationFloatWorkload_Execute", this->GetGuid());
116 RunClFunction(m_Layer, CHECK_LOCATION());
117 }
118
FreeUnusedTensors()119 void ClBatchNormalizationFloatWorkload::FreeUnusedTensors()
120 {
121 FreeTensorIfUnused(m_Mean);
122 FreeTensorIfUnused(m_Variance);
123 FreeTensorIfUnused(m_Gamma);
124 FreeTensorIfUnused(m_Beta);
125 }
126
ReplaceInputTensorHandle(ITensorHandle * tensorHandle,unsigned int slot)127 void ClBatchNormalizationFloatWorkload::ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
128 {
129 ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
130 this->m_Data.m_Inputs[slot] = tensorHandle;
131 try
132 {
133 Reconfigure();
134 }
135 catch(armnn::UnimplementedException& e)
136 {
137 // Cannot reconfigure, revert the slot back and throw the exception.
138 this->m_Data.m_Inputs[slot] = backupHandle;
139 throw e;
140 }
141 }
142
143 // Replace output tensor handle with the given TensorHandle
ReplaceOutputTensorHandle(ITensorHandle * tensorHandle,unsigned int slot)144 void ClBatchNormalizationFloatWorkload::ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
145 {
146 ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
147 this->m_Data.m_Inputs[slot] = tensorHandle;
148 try
149 {
150 Reconfigure();
151 }
152 catch(armnn::UnimplementedException& e)
153 {
154 // Cannot reconfigure, revert the slot back and throw the exception.
155 this->m_Data.m_Inputs[slot] = backupHandle;
156 throw e;
157 }
158 }
159
Reconfigure()160 void ClBatchNormalizationFloatWorkload::Reconfigure()
161 {
162 throw armnn::UnimplementedException("Reconfigure not implemented for this workload");
163 }
164
165 } //namespace armnn
166