1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonSplitterWorkload.hpp"
7
8 #include <arm_compute/runtime/NEON/functions/NESplit.h>
9
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 #include <aclCommon/ArmComputeUtils.hpp>
12 #include <armnn/utility/PolymorphicDowncast.hpp>
13 #include <armnn/backends/TensorHandle.hpp>
14 #include <neon/NeonTensorHandle.hpp>
15
16 #include "NeonWorkloadUtils.hpp"
17
18 namespace armnn
19 {
20
21 using namespace armcomputetensorutils;
22
23 namespace
24 {
CalcAclAxis(unsigned int numDimensions,unsigned int splitAxis)25 unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int splitAxis)
26 {
27 return (numDimensions - splitAxis) - 1;
28 }
29
30 } //namespace
31
NeonSplitterWorkloadValidate(const TensorInfo & input,const std::vector<std::reference_wrapper<TensorInfo>> & outputs,unsigned int splitAxis)32 arm_compute::Status NeonSplitterWorkloadValidate(const TensorInfo& input,
33 const std::vector<std::reference_wrapper<TensorInfo>>& outputs,
34 unsigned int splitAxis)
35 {
36 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
37
38 size_t numOutputs = outputs.size();
39
40 std::vector<arm_compute::TensorInfo> aclOutputs;
41 aclOutputs.reserve(numOutputs);
42
43 std::vector<arm_compute::ITensorInfo*> aclOutputPtr;
44 aclOutputPtr.reserve(numOutputs);
45
46 for (size_t i = 0u; i < outputs.size(); ++i)
47 {
48 aclOutputs.emplace_back(BuildArmComputeTensorInfo(outputs[i]));
49 aclOutputPtr.emplace_back(&aclOutputs.back());
50 }
51
52 unsigned int aclAxis = CalcAclAxis(input.GetNumDimensions(), splitAxis);
53 return arm_compute::NESplit::validate(&aclInputInfo, aclOutputPtr, aclAxis);
54 }
55
NeonSplitterWorkload(const SplitterQueueDescriptor & descriptor,const WorkloadInfo & info)56 NeonSplitterWorkload::NeonSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info)
57 : NeonBaseWorkload<SplitterQueueDescriptor>(descriptor, info)
58 {
59 // Report Profiling Details
60 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonSplitterWorkload_Construct",
61 descriptor.m_Parameters,
62 info,
63 this->GetGuid());
64
65 bool allOutputsAreSubtensors = true;
66
67 // Check that all outputs are sub-tensors
68 for (auto output : m_Data.m_Outputs)
69 {
70 if (output && !output->GetParent())
71 {
72 // Non sub-tensor input found so we need to execute the split function
73 allOutputsAreSubtensors = false;
74 break;
75 }
76 }
77
78 if (allOutputsAreSubtensors)
79 {
80 // Can skip configuring the split function since it's not executed
81 return;
82 }
83
84 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
85
86 std::vector<arm_compute::ITensor *> aclOutputs;
87 for (auto output : m_Data.m_Outputs)
88 {
89 arm_compute::ITensor& aclOutput = PolymorphicPointerDowncast<IAclTensorHandle>(output)->GetTensor();
90 aclOutputs.emplace_back(&aclOutput);
91 }
92
93 // Create the layer function
94 std::unique_ptr<arm_compute::NESplit> layer(new arm_compute::NESplit());
95
96 // Configure input and output tensors
97 std::set<unsigned int> splitAxis = ComputeSplitAxis(descriptor.m_Parameters, m_Data.m_Inputs[0]->GetShape());
98 if (splitAxis.size() != 1)
99 {
100 throw InvalidArgumentException("Cannot derive split axis from SplitterDescriptor");
101 }
102
103 unsigned int aclAxis = CalcAclAxis(descriptor.m_Parameters.GetNumDimensions(), *splitAxis.begin());
104 layer->configure(&input, aclOutputs, aclAxis);
105
106 // Prepare
107 layer->prepare();
108 m_Layer.reset(layer.release());
109 }
110
Execute() const111 void NeonSplitterWorkload::Execute() const
112 {
113 if (m_Layer)
114 {
115 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonSplitterWorkload_Execute", this->GetGuid());
116 m_Layer->run();
117 }
118 }
119
120 } //namespace armnn
121
122