xref: /aosp_15_r20/external/armnn/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp (revision 89c4ff92f2867872bb9e2354d150bf0c8c502810)
1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "ClTransposeConvolution2dWorkload.hpp"
7 
8 #include "ClWorkloadUtils.hpp"
9 
10 #include <cl/ClLayerSupport.hpp>
11 #include <cl/ClTensorHandle.hpp>
12 #include <cl/ClLayerSupport.hpp>
13 
14 #include <aclCommon/ArmComputeUtils.hpp>
15 #include <aclCommon/ArmComputeTensorUtils.hpp>
16 
17 #include <armnn/backends/TensorHandle.hpp>
18 
19 #include <arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h>
20 
21 namespace armnn
22 {
23 
24 using namespace armcomputetensorutils;
25 
ClTransposeConvolution2dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const TransposeConvolution2dDescriptor & descriptor,const TensorInfo & weights,const Optional<TensorInfo> & biases)26 arm_compute::Status ClTransposeConvolution2dWorkloadValidate(const TensorInfo& input,
27                                                              const TensorInfo& output,
28                                                              const TransposeConvolution2dDescriptor& descriptor,
29                                                              const TensorInfo& weights,
30                                                              const Optional<TensorInfo>& biases)
31 {
32     arm_compute::TensorInfo aclInputInfo   = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
33     arm_compute::TensorInfo aclOutputInfo  = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
34     arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
35 
36     arm_compute::TensorInfo aclBiasesInfo;
37     arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
38 
39     if (descriptor.m_BiasEnabled)
40     {
41         ARMNN_ASSERT(biases.has_value());
42 
43         aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
44         optionalAclBiasesInfo = &aclBiasesInfo;
45     }
46 
47     arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(descriptor);
48 
49     return arm_compute::CLDeconvolutionLayer::validate(&aclInputInfo,
50                                                        &aclWeightsInfo,
51                                                        optionalAclBiasesInfo,
52                                                        &aclOutputInfo,
53                                                        padStrideInfo);
54 }
55 
ClTransposeConvolution2dWorkload(const TransposeConvolution2dQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager,const arm_compute::CLCompileContext & clCompileContext)56 ClTransposeConvolution2dWorkload::ClTransposeConvolution2dWorkload(
57     const TransposeConvolution2dQueueDescriptor& descriptor,
58     const WorkloadInfo& info,
59     std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
60     const arm_compute::CLCompileContext& clCompileContext)
61     : ClBaseWorkload<TransposeConvolution2dQueueDescriptor>(descriptor, info)
62     , m_Layer(memoryManager)
63 {
64     // Add details for profiling output
65     WorkloadInfo detailsInfo;
66 
67     detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
68     detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
69     detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo());
70     if (descriptor.m_Parameters.m_BiasEnabled)
71     {
72         detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo());
73     }
74 
75     // Report Profiling Details
76     ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClTransposeConvolution2dWorkload_Construct",
77                                          descriptor.m_Parameters,
78                                          detailsInfo,
79                                          this->GetGuid());
80 
81     const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
82 
83     m_WeightsTensor = std::make_unique<arm_compute::CLTensor>();
84     BuildArmComputeTensor(*m_WeightsTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
85 
86     if (m_Data.m_Parameters.m_BiasEnabled)
87     {
88         m_BiasesTensor = std::make_unique<arm_compute::CLTensor>();
89         BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
90     }
91 
92     m_Data.ValidateInputsOutputs("ClTransposeConvolution2dWorkload", 1, 1);
93 
94     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
95     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
96 
97     arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
98 
99     input.info()->set_data_layout(aclDataLayout);
100     output.info()->set_data_layout(aclDataLayout);
101 
102     arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
103     {
104         ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClTransposeConvolution2dWorkload_configure");
105         m_Layer.configure(clCompileContext, &input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output,
106                           padStrideInfo);
107     }
108 
109     InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight);
110     if (m_BiasesTensor)
111     {
112         InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias);
113     }
114 
115     m_Layer.prepare();
116 
117     FreeUnusedTensors();
118 }
119 
Execute() const120 void ClTransposeConvolution2dWorkload::Execute() const
121 {
122     ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClTransposeConvolution2dWorkload_Execute", this->GetGuid());
123     RunClFunction(m_Layer, CHECK_LOCATION());
124 }
125 
FreeUnusedTensors()126 void ClTransposeConvolution2dWorkload::FreeUnusedTensors()
127 {
128     FreeTensorIfUnused(m_WeightsTensor);
129     FreeTensorIfUnused(m_BiasesTensor);
130 }
131 
132 } // namespace armnn
133