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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#include "backends/CpuTensorHandle.hpp"
#include "backends/ArmComputeTensorUtils.hpp"
#include "backends/NeonLayerSupport.hpp"
#include "NeonConvolution2dBaseWorkload.hpp"
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const TensorInfo& biases)
{
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights);
arm_compute::TensorInfo aclBiasesInfo;
arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
if (descriptor.m_BiasEnabled)
{
aclBiasesInfo = BuildArmComputeTensorInfo(biases);
optionalAclBiasesInfo = &aclBiasesInfo;
}
arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
return arm_compute::NEConvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
layerInfo);
}
template<armnn::DataType dataType>
NeonConvolution2dBaseWorkload<dataType>::NeonConvolution2dBaseWorkload(const Convolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
: TypedWorkload<Convolution2dQueueDescriptor, dataType>(descriptor, info)
{
using arm_compute::NEDirectConvolutionLayer;
using namespace armcomputetensorutils;
ValidateData();
// 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();
BuildArmComputeTensor(m_KernelTensor, m_Data.m_Weight->GetTensorInfo());
arm_compute::Tensor* optionalBiasTensor = nullptr;
if (m_Data.m_Parameters.m_BiasEnabled)
{
BuildArmComputeTensor(m_BiasTensor, m_Data.m_Bias->GetTensorInfo());
optionalBiasTensor = &m_BiasTensor;
}
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,
optionalBiasTensor,
&output,
padStrideInfo);
m_ConvolutionLayer.reset(directConvolutionLayer.release());
}
else
{
auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
convolutionLayer->configure(&input,
&m_KernelTensor,
optionalBiasTensor,
&output,
padStrideInfo);
m_ConvolutionLayer.reset(convolutionLayer.release());
}
BOOST_ASSERT(m_ConvolutionLayer);
using Type = ResolveType<dataType>;
InitialiseArmComputeTensorData(m_KernelTensor, m_Data.m_Weight->template GetConstTensor<Type>());
}
// Generate known implementations for linker
template class NeonConvolution2dBaseWorkload<DataType::Float32>;
template class NeonConvolution2dBaseWorkload<DataType::QuantisedAsymm8>;
} //namespace armnn
|
