1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonConcatWorkload.hpp"
7
8 #include "NeonWorkloadUtils.hpp"
9
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 #include <armnn/utility/PolymorphicDowncast.hpp>
12 #include <armnn/backends/TensorHandle.hpp>
13 #include <neon/NeonTensorHandle.hpp>
14
15 namespace armnn
16 {
17 using namespace armcomputetensorutils;
18
19 namespace
20 {
CalcAxis(const armnn::OriginsDescriptor & descriptor)21 size_t CalcAxis(const armnn::OriginsDescriptor& descriptor)
22 {
23 return (descriptor.GetNumDimensions() - descriptor.GetConcatAxis()) - 1;
24 }
25 } //namespace
26
NeonConcatWorkloadValidate(const std::vector<const TensorInfo * > & inputs,const TensorInfo & output,const OriginsDescriptor & descriptor)27 arm_compute::Status NeonConcatWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
28 const TensorInfo& output,
29 const OriginsDescriptor& descriptor)
30
31 {
32 std::vector<arm_compute::TensorInfo> aclInputs;
33 for (const TensorInfo* input : inputs)
34 {
35 arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW);
36 aclInputs.emplace_back(aclInputInfo);
37 }
38 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
39 std::vector<const arm_compute::ITensorInfo*> aclInputPtrs;
40 for (arm_compute::ITensorInfo& input : aclInputs)
41 {
42 aclInputPtrs.emplace_back(&input);
43 }
44
45 size_t aclAxis = CalcAxis(descriptor);
46 return arm_compute::NEConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis);
47 }
48
NeonConcatWorkload(const ConcatQueueDescriptor & descriptor,const WorkloadInfo & info)49 NeonConcatWorkload::NeonConcatWorkload(
50 const ConcatQueueDescriptor& descriptor, const WorkloadInfo& info)
51 : NeonBaseWorkload<ConcatQueueDescriptor>(descriptor, info)
52 {
53 // Report Profiling Details
54 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConcatWorkload_Construct",
55 descriptor.m_Parameters,
56 info,
57 this->GetGuid());
58
59 bool allInputsAreSubtensors = true;
60
61 // Check that all inputs are sub-tensors
62 for (auto input : descriptor.m_Inputs)
63 {
64 if (!input->GetParent())
65 {
66 // Non sub-tensor input found so we need to execute the concat function
67 allInputsAreSubtensors = false;
68 break;
69 }
70 }
71
72 if (allInputsAreSubtensors)
73 {
74 // Can skip configuring the concat function since it's not executed
75 return;
76 }
77
78 std::vector<const arm_compute::ITensor *> aclInputs;
79 for (auto input : m_Data.m_Inputs)
80 {
81 arm_compute::ITensor& aclInput = armnn::PolymorphicPointerDowncast<IAclTensorHandle>(input)->GetTensor();
82 aclInputs.emplace_back(&aclInput);
83 }
84 arm_compute::ITensor& output = armnn::PolymorphicPointerDowncast<IAclTensorHandle>(
85 m_Data.m_Outputs[0])->GetTensor();
86
87 // Create the layer function
88 m_Layer.reset(new arm_compute::NEConcatenateLayer());
89
90 // Configure input and output tensors
91 size_t aclAxis = CalcAxis(descriptor.m_Parameters);
92 m_Layer->configure(aclInputs, &output, aclAxis);
93
94 // Prepare
95 m_Layer->prepare();
96 }
97
Execute() const98 void NeonConcatWorkload::Execute() const
99 {
100 if (m_Layer)
101 {
102 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConcatWorkload_Execute", this->GetGuid());
103 m_Layer->run();
104 }
105 }
106
107 } //namespace armnn
108
109