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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
|
//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClBatchNormalizationFloatWorkload.hpp"
#include "ClWorkloadUtils.hpp"
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <armnn/backends/TensorHandle.hpp>
#include <cl/ClLayerSupport.hpp>
#include <cl/ClTensorHandle.hpp>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status 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)
{
const arm_compute::TensorInfo aclInputInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclMeanInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(mean, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclVarInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(var, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclBetaInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(beta, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclGammaInfo =
armcomputetensorutils::BuildArmComputeTensorInfo(gamma, descriptor.m_DataLayout);
const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
activationDescriptor);
return arm_compute::CLBatchNormalizationLayer::validate(&aclInputInfo,
&aclOutputInfo,
&aclMeanInfo,
&aclVarInfo,
&aclBetaInfo,
&aclGammaInfo,
descriptor.m_Eps,
activationInfo);
}
ClBatchNormalizationFloatWorkload::ClBatchNormalizationFloatWorkload(
const BatchNormalizationQueueDescriptor& descriptor,
const WorkloadInfo& info,
const arm_compute::CLCompileContext& clCompileContext)
: FloatWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
{
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchNormalizationWorkload_Construct",
descriptor.m_Parameters,
info,
this->GetGuid());
m_Mean = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo());
m_Variance = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo());
m_Gamma = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
m_Beta = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo());
m_Data.ValidateInputsOutputs("ClBatchNormalizationFloatWorkload", 1, 1);
arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(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::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClBatchNormalizationFloatWorkload_configure");
m_Layer.configure(clCompileContext,
&input,
&output,
m_Mean.get(),
m_Variance.get(),
m_Beta.get(),
m_Gamma.get(),
m_Data.m_Parameters.m_Eps,
activationInfo);
}
InitializeArmComputeClTensorData(*m_Mean, m_Data.m_Mean);
InitializeArmComputeClTensorData(*m_Variance, m_Data.m_Variance);
InitializeArmComputeClTensorData(*m_Beta, m_Data.m_Beta);
InitializeArmComputeClTensorData(*m_Gamma, m_Data.m_Gamma);
// Force Compute Library to perform the necessary copying and reshaping, after which
// delete all the input tensors that will no longer be needed
m_Layer.prepare();
FreeUnusedTensors();
}
void ClBatchNormalizationFloatWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClBatchNormalizationFloatWorkload_Execute", this->GetGuid());
RunClFunction(m_Layer, CHECK_LOCATION());
}
void ClBatchNormalizationFloatWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_Mean);
FreeTensorIfUnused(m_Variance);
FreeTensorIfUnused(m_Gamma);
FreeTensorIfUnused(m_Beta);
}
void ClBatchNormalizationFloatWorkload::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 ClBatchNormalizationFloatWorkload::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 ClBatchNormalizationFloatWorkload::Reconfigure()
{
throw armnn::UnimplementedException("Reconfigure not implemented for this workload");
}
} //namespace armnn
|
