1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #include "ClStackWorkload.hpp"
6 #include "ClWorkloadUtils.hpp"
7 #include <aclCommon/ArmComputeTensorUtils.hpp>
8 #include <armnn/utility/NumericCast.hpp>
9 #include <armnn/utility/PolymorphicDowncast.hpp>
10 #include <backendsCommon/CpuTensorHandle.hpp>
11 #include <cl/ClTensorHandle.hpp>
12 #include <cl/ClLayerSupport.hpp>
13
14 #include <arm_compute/core/Types.h>
15
16 namespace armnn
17 {
18 using namespace armcomputetensorutils;
19
20 namespace
21 {
CalcAxis(const unsigned int axis,const unsigned int inputDimensions)22 int CalcAxis(const unsigned int axis, const unsigned int inputDimensions)
23 {
24 const int intAxis = armnn::numeric_cast<int>(axis);
25 return armnn::numeric_cast<int>(inputDimensions) - intAxis;
26 }
27 } //namespace
28
ClStackWorkloadValidate(const std::vector<const TensorInfo * > & inputs,const TensorInfo & output,const StackDescriptor & descriptor)29 arm_compute::Status ClStackWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
30 const TensorInfo& output,
31 const StackDescriptor& descriptor)
32 {
33 std::vector<arm_compute::ITensorInfo*> aclInputPtrs;
34 arm_compute::TensorInfo aclInputInfo;
35 for (const TensorInfo* input : inputs)
36 {
37 aclInputInfo = BuildArmComputeTensorInfo(*input);
38 aclInputPtrs.emplace_back(&aclInputInfo);
39 }
40 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
41
42 int aclAxis = CalcAxis(descriptor.m_Axis, descriptor.m_InputShape.GetNumDimensions());
43
44 return arm_compute::CLStackLayer::validate(aclInputPtrs, aclAxis, &aclOutputInfo);
45 }
46
ClStackWorkload(const StackQueueDescriptor & descriptor,const WorkloadInfo & info)47 ClStackWorkload::ClStackWorkload(const StackQueueDescriptor& descriptor, const WorkloadInfo& info)
48 : BaseWorkload<StackQueueDescriptor>(descriptor, info)
49 {
50 std::vector<arm_compute::ICLTensor*> aclInputs;
51 for (auto input : m_Data.m_Inputs)
52 {
53 arm_compute::ICLTensor& aclInput = armnn::PolymorphicPointerDowncast<IClTensorHandle>(input)->GetTensor();
54 aclInputs.emplace_back(&aclInput);
55 }
56 arm_compute::ICLTensor& output = armnn::PolymorphicPointerDowncast<IClTensorHandle>(
57 m_Data.m_Outputs[0])->GetTensor();
58
59 m_Layer.reset(new arm_compute::CLStackLayer());
60 int aclAxis = CalcAxis(descriptor.m_Parameters.m_Axis, descriptor.m_Parameters.m_InputShape.GetNumDimensions());
61 m_Layer->configure(aclInputs, aclAxis, &output);
62 }
63
Execute() const64 void ClStackWorkload::Execute() const
65 {
66 if (m_Layer)
67 {
68 ARMNN_SCOPED_PROFILING_EVENT_CL("ClStackWorkload_Execute");
69 m_Layer->run();
70 }
71 }
72
73 } //namespace armnn