1 //
2 // Copyright © 2019 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonArgMinMaxWorkload.hpp"
7 #include "NeonWorkloadUtils.hpp"
8
9 #include <aclCommon/ArmComputeTensorUtils.hpp>
10
11 #include <armnn/backends/TensorHandle.hpp>
12
13 #include <armnn/utility/NumericCast.hpp>
14 #include <armnn/utility/PolymorphicDowncast.hpp>
15 #include <armnnUtils/TensorUtils.hpp>
16
17 #include <arm_compute/runtime/NEON/functions/NEArgMinMaxLayer.h>
18
19 namespace
20 {
CalcAclAxis(unsigned int numDimensions,unsigned int axisIndex)21 unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int axisIndex)
22 {
23 return (numDimensions - axisIndex) - 1;
24 }
25
26 } //namespace
27
28 namespace armnn
29 {
30
NeonArgMinMaxWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const ArgMinMaxDescriptor & descriptor)31 arm_compute::Status NeonArgMinMaxWorkloadValidate(const TensorInfo& input,
32 const TensorInfo& output,
33 const ArgMinMaxDescriptor& descriptor)
34 {
35 const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input);
36 const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
37
38 auto numDims = input.GetNumDimensions();
39 auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, descriptor.m_Axis);
40 int aclAxis = armnn::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis));
41
42 if (descriptor.m_Function == ArgMinMaxFunction::Max)
43 {
44 return arm_compute::NEArgMinMaxLayer::validate(&aclInput, aclAxis, &aclOutput,
45 arm_compute::ReductionOperation::ARG_IDX_MAX);
46 }
47 else
48 {
49 return arm_compute::NEArgMinMaxLayer::validate(&aclInput, aclAxis, &aclOutput,
50 arm_compute::ReductionOperation::ARG_IDX_MIN);
51 }
52 }
53
54
NeonArgMinMaxWorkload(const ArgMinMaxQueueDescriptor & descriptor,const WorkloadInfo & info)55 NeonArgMinMaxWorkload::NeonArgMinMaxWorkload(const ArgMinMaxQueueDescriptor& descriptor,
56 const WorkloadInfo& info)
57 : NeonBaseWorkload<ArgMinMaxQueueDescriptor>(descriptor, info)
58 {
59 // Report Profiling Details
60 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonArgMinMaxWorkload_Construct",
61 descriptor.m_Parameters,
62 info,
63 this->GetGuid());
64
65 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
66 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
67
68 auto numDims = info.m_InputTensorInfos[0].GetNumDimensions();
69 auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, m_Data.m_Parameters.m_Axis);
70 int aclAxis = armnn::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis));
71
72 auto layer = std::make_unique<arm_compute::NEArgMinMaxLayer>();
73
74 if (m_Data.m_Parameters.m_Function == ArgMinMaxFunction::Max)
75 {
76 layer->configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MAX);
77 }
78 else
79 {
80 layer->configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MIN);
81 }
82
83 m_ArgMinMaxLayer.reset(layer.release());
84 }
85
Execute() const86 void NeonArgMinMaxWorkload::Execute() const
87 {
88 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonArgMinMaxWorkload_Execute", this->GetGuid());
89 m_ArgMinMaxLayer->run();
90 }
91
92 } //namespace armnn
93
94