1 //
2 // Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonReduceWorkload.hpp"
7
8 #include <aclCommon/ArmComputeUtils.hpp>
9 #include <aclCommon/ArmComputeTensorUtils.hpp>
10
11 #include <neon/NeonTensorHandle.hpp>
12
13 #include "NeonWorkloadUtils.hpp"
14
15 namespace armnn
16 {
17 using namespace armcomputetensorutils;
18
NeonReduceWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const ReduceDescriptor & descriptor)19 arm_compute::Status NeonReduceWorkloadValidate(const TensorInfo& input,
20 const TensorInfo& output,
21 const ReduceDescriptor& descriptor)
22 {
23 if ( descriptor.m_vAxis.size()==1 || descriptor.m_vAxis.empty())
24 {
25 const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
26 const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
27
28 arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
29 input.GetNumDimensions(),
30 descriptor.m_vAxis);
31
32 return arm_compute::NEReductionOperation::validate(&aclInputInfo,
33 &aclOutputInfo,
34 static_cast<unsigned int>(coords[0]),
35 ConvertReductionOperationToAcl(descriptor),
36 descriptor.m_KeepDims);
37 }
38 else
39 {
40 // Validate layer if there are multiple axes.
41 arm_compute::Status status;
42 IS_MULTI_AXES_REDUCE_SUPPORTED(NeonReduceWorkloadValidate, input, descriptor, status);
43 return status;
44 }
45 }
46
NeonReduceWorkload(const ReduceQueueDescriptor & descriptor,const WorkloadInfo & info)47 NeonReduceWorkload::NeonReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info)
48 : NeonBaseWorkload<ReduceQueueDescriptor>(descriptor, info)
49 {
50 // Report Profiling Details
51 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonReduceWorkload_Construct",
52 descriptor.m_Parameters,
53 info,
54 this->GetGuid());
55
56 m_Data.ValidateInputsOutputs("NeonReduceWorkload", 1, 1);
57
58 arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
59 arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
60
61 arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(),
62 info.m_InputTensorInfos[0].GetNumDimensions(),
63 m_Data.m_Parameters.m_vAxis);
64
65 m_Layer.configure(&input,
66 &output,
67 static_cast<unsigned int>(coords[0]),
68 ConvertReductionOperationToAcl(m_Data.m_Parameters),
69 m_Data.m_Parameters.m_KeepDims);
70 }
71
Execute() const72 void NeonReduceWorkload::Execute() const
73 {
74 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonReduceWorkload_Execute", this->GetGuid());
75 m_Layer.run();
76 }
77
78 } //namespace armnn
79