1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #include "NeonStackWorkload.hpp"
6 #include "NeonWorkloadUtils.hpp"
7
8 #include <aclCommon/ArmComputeTensorUtils.hpp>
9 #include <armnn/utility/NumericCast.hpp>
10 #include <armnn/utility/PolymorphicDowncast.hpp>
11 #include <backendsCommon/CpuTensorHandle.hpp>
12 #include <neon/NeonTensorHandle.hpp>
13
14 namespace armnn
15 {
16 using namespace armcomputetensorutils;
17
18 namespace
19 {
CalcAxis(const unsigned int axis,const unsigned int inputDimensions)20 int CalcAxis(const unsigned int axis, const unsigned int inputDimensions)
21 {
22 const int intAxis = armnn::numeric_cast<int>(axis);
23 return armnn::numeric_cast<int>(inputDimensions) - intAxis;
24 }
25 } //namespace
26
NeonStackWorkloadValidate(const std::vector<const TensorInfo * > & inputs,const TensorInfo & output,const StackDescriptor & descriptor)27 arm_compute::Status NeonStackWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
28 const TensorInfo& output,
29 const StackDescriptor& descriptor)
30 {
31 std::vector<arm_compute::TensorInfo> aclInputs;
32 for (const TensorInfo* input : inputs)
33 {
34 arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW);
35 aclInputs.emplace_back(aclInputInfo);
36 }
37
38 std::vector<arm_compute::ITensorInfo*> aclInputPtrs;
39 for (arm_compute::ITensorInfo& input : aclInputs)
40 {
41 aclInputPtrs.emplace_back(&input);
42 }
43
44 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
45 int aclAxis = CalcAxis(descriptor.m_Axis, descriptor.m_InputShape.GetNumDimensions());
46 return arm_compute::NEStackLayer::validate(aclInputPtrs, aclAxis, &aclOutputInfo);
47 }
48
NeonStackWorkload(const StackQueueDescriptor & descriptor,const WorkloadInfo & info)49 NeonStackWorkload::NeonStackWorkload(const StackQueueDescriptor& descriptor, const WorkloadInfo& info)
50 : BaseWorkload<StackQueueDescriptor>(descriptor, info)
51 {
52 std::vector<arm_compute::ITensor*> aclInputs;
53 for (auto input : m_Data.m_Inputs)
54 {
55 arm_compute::ITensor& aclInput = PolymorphicPointerDowncast<IAclTensorHandle>(input)->GetTensor();
56 aclInputs.emplace_back(&aclInput);
57 }
58 arm_compute::ITensor& output = PolymorphicPointerDowncast<IAclTensorHandle>(
59 m_Data.m_Outputs[0])->GetTensor();
60
61 m_Layer.reset(new arm_compute::NEStackLayer());
62 int aclAxis = CalcAxis(descriptor.m_Parameters.m_Axis, descriptor.m_Parameters.m_InputShape.GetNumDimensions());
63 m_Layer->configure(aclInputs, aclAxis, &output);
64 }
65
Execute() const66 void NeonStackWorkload::Execute() const
67 {
68 if (m_Layer)
69 {
70 ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonStackWorkload_Execute");
71 m_Layer->run();
72 }
73 }
74
75 } //namespace armnn
76