• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //
2 // Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "NeonConvolution2dWorkload.hpp"
7 
8 #include <aclCommon/ArmComputeTensorUtils.hpp>
9 #include <aclCommon/ArmComputeUtils.hpp>
10 #include <armnn/utility/PolymorphicDowncast.hpp>
11 #include <armnn/backends/TensorHandle.hpp>
12 #include <neon/workloads/NeonWorkloadUtils.hpp>
13 
14 #include <arm_compute/runtime/NEON/functions/NEConvolutionLayer.h>
15 
16 #include <armnn/Types.hpp>
17 #include <Half.hpp>
18 
19 namespace armnn
20 {
21 
22 using namespace armcomputetensorutils;
23 
NeonConvolution2dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const Convolution2dDescriptor & descriptor,const TensorInfo & weights,const Optional<TensorInfo> & biases,bool isFastMathEnabled,const ActivationDescriptor * activationDescriptor)24 arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo& input,
25                                                       const TensorInfo& output,
26                                                       const Convolution2dDescriptor& descriptor,
27                                                       const TensorInfo& weights,
28                                                       const Optional<TensorInfo>& biases,
29                                                       bool isFastMathEnabled,
30                                                       const ActivationDescriptor* activationDescriptor)
31 {
32     const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
33     const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
34     arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
35     aclWeightsInfo.set_are_values_constant(weights.IsConstant());
36 
37     const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX,
38                                                                       descriptor.m_DilationY);
39 
40     arm_compute::TensorInfo aclBiasesInfo;
41     arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
42 
43     if (descriptor.m_BiasEnabled)
44     {
45         ARMNN_ASSERT(biases.has_value());
46         // Same for bias as weights. We don't currently support non const.
47         if (!biases.value().IsConstant())
48         {
49             return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
50                                        "ArmNN NeonConvolution2dWorkload does not support non constant bias."};
51         }
52         aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
53         aclBiasesInfo.set_are_values_constant(biases.value().IsConstant());
54         optionalAclBiasesInfo = &aclBiasesInfo;
55     }
56 
57     arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
58 
59     const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
60             activationDescriptor);
61 
62     return arm_compute::NEConvolutionLayer::validate(&aclInputInfo,
63                                                      &aclWeightsInfo,
64                                                      optionalAclBiasesInfo,
65                                                      &aclOutputInfo,
66                                                      layerInfo,
67                                                      arm_compute::WeightsInfo(),
68                                                      aclDilationInfo,
69                                                      activationInfo,
70                                                      isFastMathEnabled);
71 }
72 
NeonConvolution2dWorkload(const Convolution2dQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager,const bool isFastMathEnabled)73 NeonConvolution2dWorkload::NeonConvolution2dWorkload(
74     const Convolution2dQueueDescriptor& descriptor,
75     const WorkloadInfo& info,
76     std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
77     const bool isFastMathEnabled)
78     : NeonBaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
79 {
80     using arm_compute::NEConvolutionLayer;
81 
82     uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
83     m_Data.ValidateInputsOutputs("NeonConvolution2dWorkload", numInputs, 1);
84 
85     arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
86     arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
87 
88     arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
89     input.info()->set_data_layout(aclDataLayout);
90     output.info()->set_data_layout(aclDataLayout);
91 
92     m_KernelTensor = std::make_unique<arm_compute::Tensor>();
93     BuildArmComputeTensor(*m_KernelTensor, info.m_InputTensorInfos[1], m_Data.m_Parameters.m_DataLayout);
94     if (m_Data.m_Parameters.m_BiasEnabled)
95     {
96         m_BiasTensor = std::make_unique<arm_compute::Tensor>();
97         BuildArmComputeTensor(*m_BiasTensor, info.m_InputTensorInfos[2], m_Data.m_Parameters.m_DataLayout);
98     }
99 
100     arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
101 
102     const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX,
103                                                                       m_Data.m_Parameters.m_DilationY);
104 
105     const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
106 
107     auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
108     convolutionLayer->configure(&input,
109                                 m_KernelTensor.get(),
110                                 m_BiasTensor.get(),
111                                 &output,
112                                 padStrideInfo,
113                                 arm_compute::WeightsInfo(),
114                                 aclDilationInfo,
115                                 activationInfo,
116                                 isFastMathEnabled);
117 
118     m_ConvolutionMethod =
119         convolutionLayer->get_convolution_method(input.info(),
120                                                  m_KernelTensor->info(),
121                                                  output.info(),
122                                                  padStrideInfo,
123                                                  arm_compute::WeightsInfo(),
124                                                  aclDilationInfo,
125                                                  activationInfo,
126                                                  isFastMathEnabled);
127 
128     // Add details for profiling output
129     WorkloadInfo detailsInfo;
130 
131     detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
132     detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
133     detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]);
134     detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod));
135 
136     if (descriptor.m_Parameters.m_BiasEnabled)
137     {
138         detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]);
139     }
140 
141     // Report Profiling Details
142     ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConvolution2dWorkload_Construct",
143                                          descriptor.m_Parameters,
144                                          detailsInfo,
145                                          GetGuid());
146 
147     m_ConvolutionLayer.reset(convolutionLayer.release());
148 
149     ARMNN_ASSERT(m_ConvolutionLayer);
150     m_KernelTensorInfo = info.m_InputTensorInfos[1];
151 
152     if (m_Data.m_Parameters.m_BiasEnabled)
153     {
154         m_BiasTensorInfo = info.m_InputTensorInfos[2];
155     }
156 }
157 
Execute() const158 void NeonConvolution2dWorkload::Execute() const
159 {
160     ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvolution2dWorkload_Execute", this->GetGuid());
161     // The constant tensors may not be fully in place until the workload is Executed
162     if (!prepared)
163     {
164         InitializeArmComputeTensorData(*m_KernelTensor, m_KernelTensorInfo, m_Data.m_Inputs[1]);
165 
166         if (m_Data.m_Parameters.m_BiasEnabled)
167         {
168             InitializeArmComputeTensorData(*m_BiasTensor, m_BiasTensorInfo, m_Data.m_Inputs[2]);
169         }
170         m_ConvolutionLayer->prepare();
171         FreeTensorIfUnused(m_KernelTensor);
172         FreeTensorIfUnused(m_BiasTensor);
173         prepared = true;
174     }
175     m_ConvolutionLayer->run();
176 }
177 
GetConvolutionMethod() const178 arm_compute::ConvolutionMethod NeonConvolution2dWorkload::GetConvolutionMethod() const
179 {
180     return m_ConvolutionMethod;
181 }
182 
183 } //namespace armnn
184