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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
|
//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonConvolution2dWorkload.hpp"
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <armnn/backends/TensorHandle.hpp>
#include <neon/workloads/NeonWorkloadUtils.hpp>
#include <arm_compute/runtime/NEON/functions/NEConvolutionLayer.h>
#include <armnn/Types.hpp>
#include <Half.hpp>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases,
bool isFastMathEnabled,
const ActivationDescriptor* activationDescriptor)
{
// The implemented workload does support both const and non const
// weights. However, in the case of non const weights we'd have to call
// prepare or configure for each inference which we're not setup to do just yet.
if (!weights.IsConstant())
{
return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
"ArmNN NeonConvolution2dWorkload does not support non constant weights."};
}
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX,
descriptor.m_DilationY);
arm_compute::TensorInfo aclBiasesInfo;
arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
if (descriptor.m_BiasEnabled)
{
ARMNN_ASSERT(biases.has_value());
// Same for bias as weights. We don't currently support non const.
if (!biases.value().IsConstant())
{
return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
"ArmNN NeonConvolution2dWorkload does not support non constant bias."};
}
aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
optionalAclBiasesInfo = &aclBiasesInfo;
}
arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
activationDescriptor);
return arm_compute::NEConvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
layerInfo,
arm_compute::WeightsInfo(),
aclDilationInfo,
activationInfo,
isFastMathEnabled);
}
NeonConvolution2dWorkload::NeonConvolution2dWorkload(
const Convolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info,
std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
const bool isFastMathEnabled)
: NeonBaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
{
using arm_compute::NEConvolutionLayer;
m_Data.ValidateInputsOutputs("NeonConvolution2dWorkload", 1, 1);
arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(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);
m_KernelTensor = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_KernelTensor, m_Data.m_Weight->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
if (m_Data.m_Parameters.m_BiasEnabled)
{
m_BiasTensor = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
}
arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX,
m_Data.m_Parameters.m_DilationY);
const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
convolutionLayer->configure(&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo,
arm_compute::WeightsInfo(),
aclDilationInfo,
activationInfo,
isFastMathEnabled);
m_ConvolutionMethod =
convolutionLayer->get_convolution_method(input.info(),
m_KernelTensor->info(),
output.info(),
padStrideInfo,
arm_compute::WeightsInfo(),
aclDilationInfo,
activationInfo,
isFastMathEnabled);
// Add details for profiling output
WorkloadInfo detailsInfo;
detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo());
detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod));
if (descriptor.m_Parameters.m_BiasEnabled)
{
detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo());
}
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConvolution2dWorkload_Construct",
descriptor.m_Parameters,
detailsInfo,
this->GetGuid());
m_ConvolutionLayer.reset(convolutionLayer.release());
ARMNN_ASSERT(m_ConvolutionLayer);
InitializeArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight);
if (m_Data.m_Parameters.m_BiasEnabled)
{
InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias);
}
m_ConvolutionLayer->prepare();
FreeUnusedTensors();
}
void NeonConvolution2dWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvolution2dWorkload_Execute", this->GetGuid());
m_ConvolutionLayer->run();
}
arm_compute::ConvolutionMethod NeonConvolution2dWorkload::GetConvolutionMethod() const
{
return m_ConvolutionMethod;
}
void NeonConvolution2dWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_KernelTensor);
FreeTensorIfUnused(m_BiasTensor);
}
} //namespace armnn
|
