1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #include "ClConcatWorkload.hpp"
6 #include "ClWorkloadUtils.hpp"
7 #include <aclCommon/ArmComputeTensorUtils.hpp>
8 #include <armnn/utility/PolymorphicDowncast.hpp>
9 #include <armnn/backends/TensorHandle.hpp>
10 #include <cl/ClTensorHandle.hpp>
11 #include <cl/ClLayerSupport.hpp>
12
13 #include <arm_compute/core/Types.h>
14
15 namespace armnn
16 {
17 using namespace armcomputetensorutils;
18
19 namespace
20 {
CalcAxis(const OriginsDescriptor & descriptor)21 size_t CalcAxis(const OriginsDescriptor& descriptor)
22 {
23 return (descriptor.GetNumDimensions() - descriptor.GetConcatAxis()) - 1;
24 }
25 } //namespace
26
ClConcatWorkloadValidate(const std::vector<const TensorInfo * > & inputs,const TensorInfo & output,const OriginsDescriptor & descriptor)27 arm_compute::Status ClConcatWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
28 const TensorInfo& output,
29 const OriginsDescriptor& 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 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
38 std::vector<const arm_compute::ITensorInfo*> aclInputPtrs;
39 for (arm_compute::ITensorInfo& input : aclInputs)
40 {
41 aclInputPtrs.emplace_back(&input);
42 }
43
44 size_t aclAxis = CalcAxis(descriptor);
45 return arm_compute::CLConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis);
46 }
47
ClConcatWorkload(const ConcatQueueDescriptor & descriptor,const WorkloadInfo & info,const arm_compute::CLCompileContext & clCompileContext)48 ClConcatWorkload::ClConcatWorkload(const ConcatQueueDescriptor& descriptor,
49 const WorkloadInfo& info,
50 const arm_compute::CLCompileContext& clCompileContext)
51 : ClBaseWorkload<ConcatQueueDescriptor>(descriptor, info)
52 {
53 // Report Profiling Details
54 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClConcatWorkload_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::ICLTensor *> aclInputs;
79 for (auto input : m_Data.m_Inputs)
80 {
81 arm_compute::ICLTensor& aclInput = armnn::PolymorphicPointerDowncast<IClTensorHandle>(input)->GetTensor();
82 aclInputs.emplace_back(&aclInput);
83 }
84
85 arm_compute::ICLTensor& output =
86 armnn::PolymorphicPointerDowncast<IClTensorHandle>(m_Data.m_Outputs[0])->GetTensor();
87
88 // Create the layer function
89 auto layer = std::make_unique<arm_compute::CLConcatenateLayer>();
90
91 {
92 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConcatWorkload_configure");
93 // Configure input and output tensors
94 size_t aclAxis = CalcAxis(descriptor.m_Parameters);
95 layer->configure(clCompileContext, aclInputs, &output, aclAxis);
96 }
97
98 // Prepare
99 layer->prepare();
100 m_Layer = std::move(layer);
101 }
102
Execute() const103 void ClConcatWorkload::Execute() const
104 {
105 if (m_Layer)
106 {
107 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClConcatWorkload_Execute", this->GetGuid());
108 m_Layer->run();
109 }
110 }
111
112 } //namespace armnn
113