1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #include "NeonTransposeConvolution2dWorkload.hpp"
6
7 #include "NeonWorkloadUtils.hpp"
8
9 #include <Profiling.hpp>
10
11 #include <armnn/Types.hpp>
12 #include <armnn/utility/PolymorphicDowncast.hpp>
13
14 #include <aclCommon/ArmComputeTensorUtils.hpp>
15
16 #include <armnn/backends/TensorHandle.hpp>
17
18 #include <neon/workloads/NeonWorkloadUtils.hpp>
19
20 namespace armnn
21 {
22
23 using namespace armcomputetensorutils;
24
NeonTransposeConvolution2dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const TransposeConvolution2dDescriptor & descriptor,const TensorInfo & weights,const Optional<TensorInfo> & biases)25 arm_compute::Status NeonTransposeConvolution2dWorkloadValidate(const TensorInfo& input,
26 const TensorInfo& output,
27 const TransposeConvolution2dDescriptor& descriptor,
28 const TensorInfo& weights,
29 const Optional<TensorInfo>& biases)
30 {
31 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
32 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
33 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
34
35 arm_compute::TensorInfo aclBiasesInfo;
36 arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
37
38 if (descriptor.m_BiasEnabled)
39 {
40 ARMNN_ASSERT(biases.has_value());
41
42 aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
43 optionalAclBiasesInfo = &aclBiasesInfo;
44 }
45
46 arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
47
48 return arm_compute::NEDeconvolutionLayer::validate(&aclInputInfo,
49 &aclWeightsInfo,
50 optionalAclBiasesInfo,
51 &aclOutputInfo,
52 layerInfo);
53 }
54
NeonTransposeConvolution2dWorkload(const TransposeConvolution2dQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager)55 NeonTransposeConvolution2dWorkload::NeonTransposeConvolution2dWorkload(
56 const TransposeConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info,
57 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
58 : NeonBaseWorkload<TransposeConvolution2dQueueDescriptor>(descriptor, info)
59 {
60 m_Data.ValidateInputsOutputs("NeonTransposeConvolution2dWorkload", 1, 1);
61
62 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
63 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
64
65 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
66 input.info()->set_data_layout(aclDataLayout);
67 output.info()->set_data_layout(aclDataLayout);
68
69 m_KernelTensor = std::make_unique<arm_compute::Tensor>();
70 BuildArmComputeTensor(*m_KernelTensor, m_Data.m_Weight->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
71
72 if (m_Data.m_Parameters.m_BiasEnabled)
73 {
74 m_BiasTensor = std::make_unique<arm_compute::Tensor>();
75 BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
76 }
77
78 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
79
80 // Add details for profiling output
81 WorkloadInfo detailsInfo;
82
83 detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
84 detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
85 detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo());
86 if (descriptor.m_Parameters.m_BiasEnabled)
87 {
88 detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo());
89 }
90
91 // Report Profiling Details
92 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonTransposeConvolution2dWorkload_Construct",
93 descriptor.m_Parameters,
94 detailsInfo,
95 this->GetGuid());
96
97 m_Layer = std::make_unique<arm_compute::NEDeconvolutionLayer>(memoryManager);
98 m_Layer->configure(&input, m_KernelTensor.get(), m_BiasTensor.get(), &output, padStrideInfo);
99
100 ARMNN_ASSERT(m_Layer);
101
102 InitializeArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight);
103
104 if (m_Data.m_Parameters.m_BiasEnabled)
105 {
106 InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias);
107 }
108
109 m_Layer->prepare();
110 FreeUnusedTensors();
111 }
112
Execute() const113 void NeonTransposeConvolution2dWorkload::Execute() const
114 {
115 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonTransposeConvolution2dWorkload_Execute", this->GetGuid());
116 m_Layer->run();
117 }
118
FreeUnusedTensors()119 void NeonTransposeConvolution2dWorkload::FreeUnusedTensors()
120 {
121 FreeTensorIfUnused(m_KernelTensor);
122 FreeTensorIfUnused(m_BiasTensor);
123 }
124
125 } // namespace armnn
126