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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonTransposeConvolution2dWorkload.hpp"
#include "NeonWorkloadUtils.hpp"
#include <Profiling.hpp>
#include <armnn/Types.hpp>
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <backendsCommon/CpuTensorHandle.hpp>
#include <neon/workloads/NeonWorkloadUtils.hpp>
#include <boost/cast.hpp>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status NeonTransposeConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const TransposeConvolution2dDescriptor& 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::NEDeconvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
layerInfo);
}
NeonTransposeConvolution2dWorkload::NeonTransposeConvolution2dWorkload(
const TransposeConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info,
std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
: BaseWorkload<TransposeConvolution2dQueueDescriptor>(descriptor, info)
{
m_Data.ValidateInputsOutputs("NeonTransposeConvolution2dWorkload", 1, 1);
arm_compute::ITensor& input = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = boost::polymorphic_downcast<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);
m_Layer = std::make_unique<arm_compute::NEDeconvolutionLayer>(memoryManager);
m_Layer->configure(&input, m_KernelTensor.get(), m_BiasTensor.get(), &output, padStrideInfo);
BOOST_ASSERT(m_Layer);
InitializeArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight);
if (m_Data.m_Parameters.m_BiasEnabled)
{
InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias);
}
m_Layer->prepare();
FreeUnusedTensors();
}
void NeonTransposeConvolution2dWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonTransposeConvolution2dWorkload_Execute");
m_Layer->run();
}
void NeonTransposeConvolution2dWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_KernelTensor);
FreeTensorIfUnused(m_BiasTensor);
}
} // namespace armnn
|
