1 //
2 // Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonConvolution3dWorkload.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/NEConv3D.h>
15
16 #include <armnn/Types.hpp>
17 #include <Half.hpp>
18
19 namespace armnn
20 {
21
22 using namespace armcomputetensorutils;
23
NeonConvolution3dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const Convolution3dDescriptor & descriptor,const TensorInfo & weights,const Optional<TensorInfo> & biases,bool isFastMathEnabled,const ActivationDescriptor * activationDescriptor)24 arm_compute::Status NeonConvolution3dWorkloadValidate(const TensorInfo& input,
25 const TensorInfo& output,
26 const Convolution3dDescriptor& 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 aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
34 arm_compute::TensorInfo aclBiasesInfo;
35 arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
36 if (descriptor.m_BiasEnabled)
37 {
38 ARMNN_ASSERT(biases.has_value());
39
40 aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
41 optionalAclBiasesInfo = &aclBiasesInfo;
42 }
43 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
44
45 const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor,
46 isFastMathEnabled,
47 activationDescriptor);
48
49 return arm_compute::NEConv3D::validate(&aclInputInfo,
50 &aclWeightsInfo,
51 optionalAclBiasesInfo,
52 &aclOutputInfo,
53 aclConv3DInfo);
54 }
55
NeonConvolution3dWorkload(const Convolution3dQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager,const bool isFastMathEnabled)56 NeonConvolution3dWorkload::NeonConvolution3dWorkload(const Convolution3dQueueDescriptor& descriptor,
57 const WorkloadInfo& info,
58 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
59 const bool isFastMathEnabled)
60 : NeonBaseWorkload<Convolution3dQueueDescriptor>(descriptor, info)
61 {
62 IgnoreUnused(memoryManager);
63
64 using arm_compute::NEConv3D;
65 uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
66 m_Data.ValidateInputsOutputs("NeonConvolution3dWorkload", numInputs, 1);
67
68 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
69 arm_compute::ITensor& weights = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
70 arm_compute::ITensor* biasesPtr = nullptr;
71 if (m_Data.m_Parameters.m_BiasEnabled)
72 {
73 biasesPtr = &PolymorphicDowncast<IAclTensorHandle *>(m_Data.m_Inputs[2])->GetTensor();
74 }
75 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
76
77 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
78 input.info()->set_data_layout(aclDataLayout);
79 weights.info()->set_data_layout(aclDataLayout);
80 output.info()->set_data_layout(aclDataLayout);
81
82 const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor, isFastMathEnabled);
83
84 auto convolutionLayer = std::make_unique<arm_compute::NEConv3D>();
85 convolutionLayer->configure(&input,
86 &weights,
87 biasesPtr,
88 &output,
89 aclConv3DInfo);
90
91 // Add details for profiling output
92 WorkloadInfo detailsInfo;
93
94 detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
95 detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
96
97 // Report Profiling Details
98 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConvolution3dWorkload_Construct",
99 descriptor.m_Parameters,
100 detailsInfo,
101 this->GetGuid());
102
103 m_ConvolutionLayer.reset(convolutionLayer.release());
104
105 ARMNN_ASSERT(m_ConvolutionLayer);
106
107 m_ConvolutionLayer->prepare();
108 }
109
Execute() const110 void NeonConvolution3dWorkload::Execute() const
111 {
112 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvolution3dWorkload_Execute", this->GetGuid());
113 m_ConvolutionLayer->run();
114 }
115
116 } //namespace armnn
117