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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonConvolution2dWorkload.hpp"
#include <backends/CpuTensorHandle.hpp>
#include <backends/aclCommon/ArmComputeTensorUtils.hpp>
#include <backends/neon/NeonLayerSupport.hpp>
#include <armnn/Types.hpp>
#include <armnnUtils/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)
{
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);
arm_compute::TensorInfo aclBiasesInfo;
arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
if (descriptor.m_BiasEnabled)
{
BOOST_ASSERT(biases.has_value());
aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
optionalAclBiasesInfo = &aclBiasesInfo;
}
arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
return arm_compute::NEConvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
layerInfo);
}
NeonConvolution2dWorkload::NeonConvolution2dWorkload(
const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info,
std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
: BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
{
using arm_compute::NEDirectConvolutionLayer;
m_Data.ValidateInputsOutputs("NeonConvolution2dWorkload", 1, 1);
// todo: check tensor shapes match.
arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(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(m_Data.m_Parameters.m_StrideX,
m_Data.m_Parameters.m_StrideY,
m_Data.m_Parameters.m_PadLeft,
m_Data.m_Parameters.m_PadRight,
m_Data.m_Parameters.m_PadTop,
m_Data.m_Parameters.m_PadBottom,
arm_compute::DimensionRoundingType::FLOOR);
const bool preferDirectConvolution =
IsNeonDirectConvolutionPreferred(m_Data.m_Weight->GetTensorInfo(),
m_Data.m_Parameters);
if (preferDirectConvolution)
{
auto directConvolutionLayer = std::make_unique<arm_compute::NEDirectConvolutionLayer>(memoryManager);
directConvolutionLayer->configure(&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo);
m_ConvolutionLayer.reset(directConvolutionLayer.release());
}
else
{
auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
convolutionLayer->configure(&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo);
m_ConvolutionLayer.reset(convolutionLayer.release());
}
BOOST_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("NeonConvolution2dWorkload_Execute");
m_ConvolutionLayer->run();
}
void NeonConvolution2dWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_KernelTensor);
FreeTensorIfUnused(m_BiasTensor);
}
} //namespace armnn
|
