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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClTransposeConvolution2dWorkload.hpp"
#include "ClWorkloadUtils.hpp"
#include <cl/ClLayerSupport.hpp>
#include <cl/ClTensorHandle.hpp>
#include <cl/ClLayerSupport.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <backendsCommon/TensorHandle.hpp>
#include <arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status ClTransposeConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const TransposeConvolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases)
{
arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
arm_compute::TensorInfo aclBiasesInfo;
arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
if (descriptor.m_BiasEnabled)
{
ARMNN_ASSERT(biases.has_value());
aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
optionalAclBiasesInfo = &aclBiasesInfo;
}
arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(descriptor);
return arm_compute::CLDeconvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
padStrideInfo);
}
ClTransposeConvolution2dWorkload::ClTransposeConvolution2dWorkload(
const TransposeConvolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info,
std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
const arm_compute::CLCompileContext& clCompileContext) :
BaseWorkload<TransposeConvolution2dQueueDescriptor>(descriptor, info),
m_Layer(memoryManager)
{
// 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());
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("ClTransposeConvolutionWorkload_Construct",
descriptor.m_Parameters,
detailsInfo,
this->GetGuid());
const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
m_WeightsTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_WeightsTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
if (m_Data.m_Parameters.m_BiasEnabled)
{
m_BiasesTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
}
m_Data.ValidateInputsOutputs("ClTransposeConvolution2dWorkload", 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);
arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
m_Layer.configure(clCompileContext, &input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, padStrideInfo);
InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight);
if (m_BiasesTensor)
{
InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias);
}
m_Layer.prepare();
FreeUnusedTensors();
}
void ClTransposeConvolution2dWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClTransposeConvolution2dWorkload_Execute", this->GetGuid());
RunClFunction(m_Layer, CHECK_LOCATION());
}
void ClTransposeConvolution2dWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_WeightsTensor);
FreeTensorIfUnused(m_BiasesTensor);
}
} // namespace armnn
|
